A survey of autonomous monitoring systems in mental health

Smartphones and personal sensing technologies have made collecting data continuously and in real time feasible. The promise of pervasive sensing technologies in the realm of mental health has recently garnered increased attention. Using Artificial Intelligence methods, it is possible to forecast a person's emotional state based on contextual information such as their current location, movement patterns, and so on. As a result, conditions like anxiety, stress, depression, and others might be tracked automatically and in real‐time. The objective of this research was to survey the state‐of‐the‐art autonomous psychological health monitoring (APHM) approaches, including those that make use of sensor data, virtual chatbot communication, and artificial intelligence methods like Machine learning and deep learning algorithms. We discussed the main processing phases of APHM from the sensing layer to the application layer and an observation taxonomy deals with various observation devices, observation duration, and phenomena related to APHM. Our goal in this study includes research works pertaining to working of APHM to predict the various mental disorders and difficulties encountered by researchers working in this sector and potential application for future clinical use highlighted.

generalized anxiety disorder (Wang et al., 2023), who studied the behavioral effects of psychosis and depression, found that patients with both diseases participated in much less motor activity than controls, as evaluated by actigraphs.Zakariah and Alotaibi investigated whether or not motor activity recording with actigraphy could be used to diagnose unipolar and bipolar depression (Zakariah & Alotaibi, 2023).
Our understanding of mental health treatment has also been bolstered by studies on the effectiveness of online and other digital interventions.Our team has completed a comprehensive analysis of studies pertaining to APHM (Autonomous Psychological Health Monitoring) methods.Wearbles, ambient sensors, smartphones, virtual agents (sources of sensing data from users) and, more importantly, AI techniques, which are used to map the sensing data to the psychological behaviors in APHM.Anxiety disorders and mood disorders were ultimate prediction from those context data based on the clinical experts intervention.Technology-enabled, autonomous mental health monitoring (multimodal sensing and various AI methodologies) are analyzed in this study.The primary results of this study are • an outline of the internal workings of APHM systems; • taxonomy for observing factors related to APHM systems; • applications of APHM systems; and • research opportunities and challenges in APHM systems.
In light of recent developments in APHM systems, we present a summary of current research focusing on analyzing human behaviors to predict the psychological disorders.Our mission is to describe the behaviors gathered by different sensing methods and to highlight promising new avenues for study in the field of behavioral sciences in this part.The inductive character of APHM systems is discussed in Section 2. Section 3 discusses about the novel taxonomy, which includes various observation categorizations of information-gathering devices, duration of observation, and observation phenomenons.Section 4 describes the AI methods in place to transform hope (sensed data form various sources) into reality (psychological disorders), various software and tools employed and measures for evaluating efficiency of AI learning methods are discussed.Applications of the APHM systems are elaborated in detail Section 5. Section 6 addresses the difficulties and potential benefits of deducing mental health issues from APHM systems.Section 6 concludes by presenting our results and the most current development.

| WORK FLOW OF AUTONOMOUS PSYCHOLOGICAL HEALTH MONITORING APHM SYSTEM
APHM systems are monitoring systems that can operate independently without human intervention.These systems commonly use context based information from sensing devices and artificial intelligence (AI) to analyze data and make choices (Tielman et al., 2019).Figure 1 provides attributes of an autonomous health monitoring system, such as realtime monitoring (sensing layer), and the system should be able to analyze the data it collects to detect patterns and abnormalities that could indicate potential health issues.Typically, after the sensing layer comes the network layer, which is responsible for data collection and storage.It can be made up of wired or wireless networks such as Bluetooth and Wi-Fi.The raw sensor data is transformed into digital forms in the analysis layer, where artificial intelligence algorithms and other advanced analytic approaches are housed.In a typical layered architecture, the application layer is the very last one.The application layer's activity categorization step will use the retrieved attributes to assign them to one of several groups.When the system detects a problem that needs attention, the application layer is designed to send alerts and notifications to the user, caregiver, or healthcare provider.It also provides remote access to the data it collects, so healthcare providers and caregivers can monitor the user's health from anywhere.

| Sensing layer
The sensor layer is essential to creating a system for monitoring psychological health, as it tracks environmental and physiological changes in the factors.Indoor environmental factors, such as humidity and temperature, may also be measured with various on-object sensors, such as environmental sensors (smart bed, smart chair, smart light, smart TV).Physiological sensors like a spirometer, temperature sensor, blood pressure sensor, and inertial sensors like an accelerometer, gyroscope, microphone via Wi-Fi and Bluetooth connectivity can be used to monitor a person's vital signs.To identify emotional and behavioral characteristics, an APHM may be designed using wearables, ambient sensors, virtual chatbots and physiological questionnaires via smartphones as shown in Table 1.Longitudinal and quantitative research methods are frequently used to collect the information needed to develop automated prediction models and monitor changes in a given condition over time.In the past, researchers have relied on well-established self-report psychiatric surveys like the Young Mania Rating Scale (YMRS) and Hamilton Depression Scale (HDS) (Gopalakrishnan, Venkataraman, Gururajan, Zhou, & Genrich, 2022;Shein et al., 2022).This method was used in tandem with wearables and smartphone applications that made use of sensors to ascertain whether the sensor-derived measures were in line with the claimed circumstances and to create causal links among the sensor-based automatedreport and the self-report.A patient's vitals can be monitored with the help of a chatbot as well.In the realm of mental health, chatbots are being used as individual medical aides to encourage psychological well-being and emotional wellness status updates before, during, and after medical intervention (Cr et al., 2020).They are also useful for detecting psychological symptoms and habits such as level of exercise, sleep schedule, and time spent on social media (Abashev et al., 2017).Constant collaboration across medical and computer science disciplines is essential for researching wearable and mobile monitoring devices.Preliminary studies are a crucial stage for these types of investigations.They involve obtaining ethical approval and user consent while meeting technical and hardware/software specifications (Riegler et al., 2016).Intrusiveness and privacy concerns play into deciding how many and what kinds of sensors to deploy, as well as technical considerations such as power consumption and sampling rates (Harari et al., 2016).The sampling rate is the pace at which data is gathered and is influenced by battery life.Higher sample rates result in more granular pattern data but are a drain on battery life (Harari et al., 2016).

| Network layer
The purpose of the network layer of an Autonomous Psychological Health Monitoring (APHM) system is to connect all sensors used in the sensory layer.The network layer allows for information gathered by wearables like watches, wristbands and virtual chatbots to be transmitted via Bluetooth, Wi-Fi, Zigbee, and cellular networks.All data, whether transferred to a secure storage platform built to handle secret data or the onboard storage of a smartphone, should be encrypted before transmission or storage to safeguard the anonymity of users (University of Oslo, 2018).

| Analysis layer
The analysis layer of an APHM system may be used to examine physiological parameters such as electrocardiogram, electro-determal activity or electroencephalogram as well as behavioral and environmental parameters such as movement patterns, social interactions, voice patterns and sleep duration, among others shown in Table 1.For performing this deduction, various processes are involved, such as Data labeling, Data preprocessing and segmentation, and deduction.

| Data labeling
Data labeling or tagging involves linking sensor data with a relevant ground truth state.This step is crucial to train the final algorithmic prediction model successfully.This can be achieved in a variety of ways, such as through periodic inperson or over-the-phone evaluations by a clinician (Faurholt-Jepsen, Vinberg, Frost, Debel, Margrethe Christensen, & Bardram, 2016;Garcia-Ceja et al., 2016) or by having a mobile app prompt users for updates at regular intervals (Ostermaier et al., 2011;Zenonos et al., 2016).

| Preparing data and extracting attributes
Exploratory data analysis and preprocessing following data collection help with data visualization, interpretation, and outlier finding.During preprocessing, filters and adjustments can be made to the raw data to get rid of extraneous information and reduce noise and outliers.Dimensionality reduction techniques used in preprocessing include principal component analysis (PCA) (Gower, 1966) and multidimensional scaling (Maxhuni, Osmani, et al., 2016).Once the raw sensor data is collected, attribute extraction is used to construct attribute vectors using word embedding (bigrams, trigrams, Bow, and n-grams).Artificial intelligence algorithms cannot function without attribute vectors representing the raw data.Minimum, maximum, root mean square, skewness, Mean, standard deviation, kurtosis, power spectrum density, correlation coefficient, and energy are often extracted parameters for determining psychological states (Garcia-Ceja, Riegler, Jakobsen, et al., 2018;Lu et al., 2012).

| Deducing
The necessary data for depression prediction can be derived using any artificial intelligence method, including Machine learning and deep learning algorithms.Supervised, unsupervised, semi-supervised, transfer, and reinforcement learning (s) are only some of the machine learning methods that can be used to sift through mountains of data in pursuit of a reliable prediction model.The final prediction models are often the result of a collaborative effort between several different algorithms.Some training approaches, such as user-dependent and user-independent (generic) models, are distinct from one another.Since they are designed to capture each user's unique behavior, user-dependent models yield superior outcomes, but they take a massive amount of data to train.Conversely, user-independent models can be trained without any information about the intended user, albeit they may perform better with atypical users.Hybrid models include the benefits of both types of models, as claimed by Lu et al. (Xu et al., 2015).However, other researchers have claimed that user-dependent models are more effective in identifying signs of stress (Zenonos et al., 2016).As previously said, psychiatric disease indices result from combining raw sensor data and analytic data platforms.When it comes to testing and training machine learning models, a wide selection of software tools and libraries are available.Anxiety may be diagnosed using the Weka (Grunerbl et al., 2014) software, while manic-depressive illness can be diagnosed using the scikit-learn package (Barnett et al., 2019).InSTIL (Intelligent Sensing to Inform and Learn) is yet another digital phenotyping instrument.Researchers use the technology to capture sensor data from consumer devices, including passive and active sensor signals.Remote Assessment of Disease and Relapse (RADAR)-the based platform was designed to assist large-scale data gathering and integration with remote monitoring activities (Ranjan et al., 2019).These platforms stress the privacy and security of their customers in the design of their services, which include remote data collection on a massive scale, research administration, and realtime visualizations.Machine learning toolboxes in Matlab are one example; other methods can be employed depending on the specifics of the application.

| Application layer
The application layer of the Autonomous Psychological Health Monitoring (APHM) system includes remote monitoring, fall detection in older adults, emotion monitoring, emergency prediction, and well-being tracking.The importance of remote Psychological health monitoring cannot be overstated for those living with Psychological illness, their loved ones, and their medical professionals.While a patient's physiological and behavioral patterns might be indicative of monitoring the Psychological state from remote (Boopathi, 2023), well-being tracking (Hurling et al., 2007)and emotion prediction (Calvo & D'Mello, 2010), elderly alone people's fall detection, and emergency alert (Sharifi Kia et al., 2023), keeping track of this information can help doctors and carers make more informed decisions about how to intervene and diagnose (Zhong et al., 2023).Depression and sad mood states associated with bipolar disorder can be detected using data from microphones, accelerometers, Wi-Fi, cameras, GPS, and mobile phone sensors (Barnett et al., 2019;Lane et al., 2010;Servia-Rodríguez et al., 2017).Stress and physiological signs can be identified using data from EDA and HR sensors.
Sleep, activity, and the rate and discourse pattern are all parameters for tracking schizophrenia (Barnett et al., 2019;Diraco et al., 2023;Servia-Rodríguez et al., 2017).Finally, cognitive support systems frequently use the microphone, light, acceleration, temperature, and digital compass sensors (Lane et al., 2010;Stankevich et al., 2012).Psychological health monitoring systems facilitate symptom tracking in the home and the hospital.Compared to traditional questionnaires or manual testing, these systems provide more specific information for physicians and carers, which may improve self-management of health, save healthcare costs, and help prevent individualized adverse outcomes.When used to offer ambient supported living, especially to older adults, these devices serve as a crucial emergency system for keeping tabs on their health and spotting any signs of unusual behavior.

| OBSERVATION-TAXONOMY
The factors related to observation in the Autonomous Psychological Health Monitoring (APHM) system are depicted in Figure 2, separated into four levels: observation study, observation phenomena, observation duration, and observation devices.As a result, each group includes subcategories of its own, such as correlation, detection, and prediction, which are the three sub types of observation study in APHM systems.Observation to predict mental health issues, both shortterm and long-term duration is possible.Finally, observation equipment can be worn or mounted outside in the environment, and physiological deduction phenomena are limited to anxiety, Bipolar defective disorder and depression.

| Observation study
Based on the previous research, the APHM system was devised with different types of study based on the primary goal of the research question.Others aim to create models that may be used to detect and predict a range of psychological states, while others seek correlation between predictor variables and the psychological state.Studies of correlation, detection, and prediction are discussed more below.Researchers conducting association studies look for connections between the collected data and the participants' mental health.These connections are utilized to accomplish things like comparing various teams.Detection is the process of predicting an existing state from a set of inputs.Predicting is the process of forecasting an outcome using historical data as input.When forecasting an outcome based on a set of given conditions, one uses the term "prediction" (Hastie et al., 2009).In this case, the term "prediction" can mean either "finding" or "making a forecast."Table 2 shows various research works under these three categories.

| Correlation
In these cases, relations between one or more input variables and Psychological well-being are observed.Participants are typically split into a "control" group and a "diagnosis" group for scientific studies.The combination of yet to be diagnosed and normal people are segregated into controlled groups and diagnosis groups of people belonging to condition of interest.Various wearable sensors, namely eye sensors, virtual reality (VR) headsets, questionnaires, accelerometers, Pupil rentia reflectors and head trackers, are used for predicting the phobia of social situations, tiredness, depression and anxiety as shown in Table 2. To extract the association between stress level and interactions the communication sequence through the microphone were found.About 90% of the observations are made correctly.It is also used to predict the variation between mood swings.Linear regression models (Diraco et al., 2023), correlation analysis (Gower, 1966), ANalysis Of VAriance (ANOVA) (Gower, 1966) and t-tests (Gower, 1966), and other statistical methods are used to detect associations and differences.

| Detection
The purpose of these works is to use real-time and historical data obtained via multiple sensory modalities to detect/ recognize a person's current Psychological state.When detecting fraud, statistical and machine-learning techniques are widely applied.To create these models for anomaly detection in the mental state, researchers use previously acquired and labeled data for training.Instead of manually discovering patterns from predictor variables, detection studies frequently employ segmentation techniques like Decision trees, Naive Bayes, and others.By observing how people use touch displays, they were able to detect the presence of pressure.Using touch variables like strength and length, they constructed a J48 decision tree classifier (Maxhuni, Osmani, et al., 2016) that achieves 78% accuracy in distinguishing between stressful and nonstressed touches.Grunerbl et al. used a Naive Bayes classifier to correctly categorize the euthymic states of bipolar illness patients from their phone call logs and audio captured by mobile microphones, achieving an accuracy rate of 76% (Hurling et al., 2007).F I G U R E 2 Observation taxonomy.

| Prediction
Time series analysis focuses on determining a frame of mind before it manifests or its signs impede a person's ability to perform.Anticipating is more complex than discovering overall, but it has more possibilities in the medical sector because actions can be carried out faster, permitting faster intervention and more successful therapies.Sirtola et al.Seventy percent of depressed people were properly identified using Random Forest.
using a quadratic discriminant classifier, (Siirtola et al., 2018) were able to anticipate migraine attacks the night before they occurred.Mormann et al. (2005) investigated using electroencephalogram (EEG) signals to predict epileptic seizures.They put their ability to discern between the preictal and interictal stages to the test to examine different metrics for seizure prediction.The preictal period occurs before a seizure, whereas the interictal stage occurs between seizures (Epilepsy Foundation, n.d.).Predictive systems are preferable to detection systems because they prompt quick responses, allowing users and caregivers to lessen or avoid uncomfortable conditions.

| Observation duration
The Perception Length is the time taken to observe the factors related to Psychological well-being.It is widely classified into short-term and long-term.The changes between distinctive Psychological states can happen in a brief span of time.In contrast, in an anxiety mess, a few shifts from tense to a relaxed might occur during the same day.Additionally, in bipolar clutters, state's moves can take a few days.Based on the circumstance, a few works collect sensor information for two to three days, while others require a few months.The short-term consideration can be 45 min or more than an hour has been carried within the paper (Seeger et al., 2011) to foresee uneasiness disorder.Slater et al. (1999), which contains three open talks which happen in a brief length of time is considered to address issues including confronting a gathering of people of avatars.
A few Psychological clutters can be measured legitimately, as it were, on the off chance that it was examined for a long time.Bipolar clutter could be related to disposition swings changing from depressive lows to hyper highs.The bipolar disorder's correct cause is unknown since it incorporates the perceptions made in genetics play a part, so it has to be dissected in the long term (Faurholt-Jepsen, Vinberg, Frost, Debel, Margrethe Christensen, Bardram, & Kessing, 2016).Smartphone sensors, including the mouthpiece, screen state, cell tower ids, SMS, and others, for 12 weeks to forecast the relationships between bipolar conditions and sensor estimations.Machine learning was used in a few articles to predict bipolar states with an accuracy of 76%.Several studies have shown that the circadian clock has an impact on Cardiovascular Disease (CVD) risk factors like heart rate and so on.Different components, such as skin conductance and heart rate variation, predict the state of the apprehensive framework, and the anxious framework contributes to the proper functioning of the circadian rhythm (Dibner et al., 2010).Table 2 provides further instances with supporting information.

| Observation device
The Observation device utilized in the APHM system for detecting physiological parameters for observing frameworks like inbuilt smartphone sensors, ambient sensors, wearable and Virtual chatbots are taken into consideration.In any case, the Psychological state can't be precisely measured utilizing the sensors but can be portrayed as detecting conduct.The circadian rhythm, for example, is determined by the light-dark cycle, rest-activity patterns that affect the sleep-wake process, and hormone secretion in humans.Changes in Circadian rhythms are a common indication of bipolar clutter; disturbed organic rhythms or unsettling influence within the circadian rhythms are considered common (Alloy et al., 2017;Panda et al., 2002).Several things have detailed that the circadian clock impacts CVD hazard components such as heart rate variability.The different attributes, such as skin conductance and heart rate fluctuation, anticipate the apprehensive framework's state.The anxious framework boosts the appropriate working of the circadian rhythm (Dibner et al., 2010;Gopalakrishnan, Venkataraman, Gururajan, Zhou, & Genrich, 2022).

| Inbuilt smart phone sensors
In this part of the review, a few studies examined how health apps make use of the sensors that are already present in smartphones.
• Microphone: Mics can be utilized to investigate inaudible emotions through wireless interaction as well.The Sound recordings within the acoustic environment give the data social intelligence and, as day-by-day exercises, behaviors captured from smartphone information (Ciman et al., 2015;Giddens et al., 2013).• Camera: The camera on a smartphone can be used to provide information about quiet pictures and recordings.The camera zooms on the images, and the data stored in the phone is taken into handling and making various forecasts.
Later investigation says that the determination of the camera surpasses distinguishing blood cell and microorganism morphology that empowers a programmed assessment.Teledermatology is a more prominent example of using portable cameras in healthcare administrations, where the specialist uses silent skin photographs to make a decision.Using unusual devices other than the phone can expand the circle of portable gadget camera applications.It is possible, for example, to create a light magnifying lens using smartphones (Lane et al., 2010;Stankevich et al., 2012).• Magnetometer and GPS: The smartphone is fitted with a magnetometer, something else known as a compass.With its capacity to sense attractive areas, this gadget identifies compass heading relative to the Earth's attractive north shaft.It determines the location of your phone using GPS (Diraco et al., 2023).Another type of sensor in your portable device is GPS, and the area is defined by satellites (Barnett et al., 2019;Lane et al., 2010;Servia-Rodríguez et al., 2017).• Gyroscope: Devices are turned in any heading and can be distinguished by a three-axis whirligig.Utilizing rotational constraints, it measures precise speed around three tomahawks.The outright introduction of your phone, speaking to as the points yaw, pitch, and roll, is identified by a combination of the accelerometer, compass, and whirligig.It decides your phone's area (Barnett et al., 2019;Diraco et al., 2023;Lane et al., 2010;Servia-Rodríguez et al., 2017).• Accelerometer: A three-axis accelerometer in your smartphone reports how quickly your phone moves in any given direct course.The accelerometer can identify gravity as an inactive speeding up and energetic, increasing speed connected to the phone (Kwapisz et al., 2010).Different sorts of Micro-Electro-Mechanical System (MEMS) accelerometer equipment are accessible, such as infinitesimal piezoelectric gems that alter voltage beneath stretch when vibrations happen or differential capacitance caused by the development of a silicon structure-the magnetometer, GPS, spinner and accelerometer on your phone from the culmination route framework (Lane et al., 2010).A few pedometers can, moreover, calculate an approximate number of burned calories.Even though pedometer like sensors can show an exceptionally valuable for physical action following, they are centered only on step tallying and don't consider another day-by-day exercise, such as strolling or running (Barnett et al., 2019;Diraco et al., 2023;Servia-Rodríguez et al., 2017).• Proximity sensor: It comprises an infrared Driven and an IR light finder.The nearness sensor identifies how near the phone is to an exterior protest, like close-by Devices.This detection is done to diminish show control utilization when you're on a call by turning off the Liquid Crystal Display (LCD) backdrop illumination.It, too, impairs the touch screen to dodge accidental touches by the cheek (Barnett et al., 2019;Diraco et al., 2023;Lane et al., 2010).• Barometer: More progressed smartphones have a chip that can identify air weight.But to utilize it, the phone should drag down neighborhood climate information for an authoritative figure on barometric weight.The interior building conditions, such as warming or air-conditioning streams, can influence the sensor's precision.These indicators work best with other devices like GPS and Wi-Fi (Barnett et al., 2019;Diraco et al., 2023).

| Wearable
Wearable sensors adaptively detect devices that customers can wear during their daily routine.Actigraphy Devices and Savvy watches/clips/bands/bracelets/sunglasses are examples of wearable sensors.
• Actigraphy devices: Actigraphy is a wrist-worn sensor-based approach for measuring human activity levels.These devices are commonly used to measure the screen rest.Two primary sleep methods, are REM (Rapid Eye Movement), and non-REM (non-Rapid Eye Movement).Each would be associated with different neural activity and cellular function.During a typical night, every individual needs to cycle across all stages of non-REM and REM sleeps numerous times, with lengthier, deeper REM bouts occurring nearer to daylight (Sadeh & Acebo, 2002).Bipolar disorder is supposed to be diagnosed through sleep.• Smartwatches/bands/bracelets: As compression of sensors, they are integrated into things like bracelets and watches.
Because these devices are in continual touch with the users' skin, sensor measurements such as skin reaction, breathing rate, warmth, and so on may be acquired.Integrating sensor information from cellphones with expert observers makes it feasible to gather more nuanced data about users' behavior (Garcia-Ceja, Riegler, Nordgreen, et al., 2018).

| Ambient sensors
Ambient sensors, or encompassing sensors, in some circumstances, are often inserted into the environment and fixed in place.They only sometimes need to be in contact with the customer, which can be beneficial because the client isn't wearing any devices.These sensors include but are not limited to, video cameras, high-quality amplifiers, and depth vision cameras (Garcia-Ceja, Riegler, Nordgreen, et al., 2018).Studying a creative environment with just a few sensors in one place is possible.Sensor data can be used to learn about the environment and tailor support, goods, and advice to locals.It may be a single bedroom, an entire apartment, or an entire city block.Sensor technology that may be easily installed and used in a wide variety of places in the homes such as Entryways, pantries, and refrigerators.They used to be able to distinguish between daily activities and physical exercises.

| Virtual agent chat
Chatbots can be integrated into numerous systems, such as mobile apps, websites, SMS texting, intelligent technologies, and virtual reality.Chatbots can employ either simple rule-based models like ELIZA or more complex AI models that combine NLP and machine learning (D'Alfonso, 2020).Rule-based models like ELIZA exist.The level of sophistication with which chatbots communicate with humans can likewise vary widely.Users can interact with chatbots in several ways, most commonly through text-based or voice-enabled interactions (Ivanovic & Semnic, 2018).This occurs when the chatbot digests the user's conversation.Most user input is textual, either as free-form text or as a set of multiplechoice options; however, the chatbot's output may be in the form of text, speech, or visuals.The user can communicate with the chatbot using open-ended text or multiple-choice questions (Shum et al., 2018).Dementia, melancholy, stress, substance misuse, anxiety disorders, suicide, and post-traumatic stress disorder (PTSD) are just some of the mental health concerns that have been identified or tested for utilizing chatbots.The customer has the same level of interaction with the chatbot as they would with a real person.The user responds to a series of questions designed to help the system diagnose their ailment, suggest a treatment plan, or do both (Ly et al., 2017).A survey of psychologists and psychiatrists indicated that 51% viewed the use of chatbots for diagnostic purposes to be harmful (Sweeney et al., 2021).However, AI diagnostics can help identify those at risk, opening the door to early intervention and reducing long-term danger (Balaskas et al., 2021).

| Observation phenomena
Psychological well-being ailment could be a general term for a bunch of issues that will affect a person's practices, thoughts, views, and sentiments.Psychological sickness can influence working and individual connections.The foremost common Psychological health issues and sicknesses are Uneasiness, clutter, Bipolar disorder, and Depression.

| Anxiety disorders
Anxiety is a negative emotion marked by apprehensiveness, worry, and even physiological manifestations like a rise in blood pressure.Anxiety sufferers often struggle with intrusive, uncontrollable thoughts and worries.Anxiety can be predicted with the help of passive sensing data collected from smartphones.Data from a smartphone's sensors, such as its microphone, GPS, camera, accelerometer, and Wi-Fi, is used to learn about the user's actions and the surrounding environment.The model is trained with machine learning using data from the user's smartphone and their own anxiety levels as inputs.Using passive sensing data, the article demonstrates how to accurately identify people who suffer from high levels of anxiety (Fukazawa et al., 2019).Anxiety can be tracked and controlled in a nonintrusive manner with the use of a smartphone.Specifically, previous research has found that (1) People who experience higher levels of worry are linked to having a lower likelihood of unlocking their phones (Rozgonjuk et al., 2018), (2) People who experience higher levels of anxiety are linked to having a smaller range of travel experiences (Boukhechba et al., 2018), and (3) People who suffer from higher levels of anxiety have been shown to spend less time in religious or contemplative settings.These passively obtained sensor data can be used with machine learning to predict the degree of trait social anxiety (r = 0.7, 85% accuracy) (Boukhechba et al., 2018;Jacobson et al., 2020) and depression (r = 0.6) (Jacobson & O'Cleirigh, 2019).These points to the possibility that the quantity of longitudinal data received passively from cellphones and wearable sensors may have a direct relationship to the intensity of anxiety symptoms.

| Bipolar affective disorder
Bipolar disorder is a psychiatric ailment that was formerly known as manic depression.It is distinguished from manic depression by the significant shifts in a person's mood, energy, activity, and focus that it causes.Because of these shifts, it's possible that our typical activities will become more difficult.Accelerometry, current place of residence, or combination of accelerometry and geolocation information produced clinical state (depression/mania) recognition accuracy ranging from 72% to 81% based on 12 patients who were studied for 12 weeks, and state change detection with 96% precision and 94% recall (Grunerbl et al., 2014).When phone call features and paralinguistic data were included, the accuracy of mood recognition increased to 76%, while the precision increased to 97% and the recall increased to 92% for state change detection (Grünerbl et al., 2015).An 18-patient 5-month study found that a smartphone app can detect stress and mood (Alvarez-Lozano et al., 2014).Social and entertainment applications reduced stress and irritability.GPS characteristics, initially established for depression, may also detect bipolar depressive episodes (Saeb, Lattie, et al., 2016).Entropy and circadian rhythm continue to predict depression severity in this sample (De Vos et al., 2016).These features also separated depressed states from non-depressed states with an accuracy of 85%.Attributes across diagnoses may be helpful for comparing similar states, and they distinguished depressed ones from non-depressed ones.

| Depression
Depression sensing has the potential to streamline the process of identifying disease-related behaviors and provide novel insights that could improve future treatments.The study of GPS depression is one instance.Using data collected over the course of 2 weeks from 28 subjects, the first study found evidence linking certain GPS-derived position metrics to depression (Saeb, Lonini, et al., 2016).Location entropy (the variety in time spent in different sites) was correlated with depression, while total number of locations visited was not.More concentrated time spent in a few places was linked to higher depression risk, while more evenly distributed time was related with lower depression scores.Interruptions in periodicity, or the circadian pattern of travel through geographical locations, were associated with more severe depressive symptoms, suggesting a strong correlation between periodicity and depression.These results were confirmed on the Student Life dataset (Saeb, Lonini, et al., 2016).Using somewhat different methods, a third study found that GPS characteristics were useful for estimating levels of depression (Canzian & Musolesi, 2015).Depression and mobility has been the subject of research.On days when social expectations do not drive behavior, such as weekends and holidays, the relationship between GPS characteristics and depression is higher (Saeb, Lonini, et al., 2016).This demonstrates that separating times when an individual's actions are more within their control from times when they are not may reveal characteristics that enhance model accuracy.A lack of mobility may be an early warning indication of sadness since GPS attributes can predict depression weeks in advance, but the association between depression and subsequent GPS attributes soon fades.

| METHODS FOR TRANSFORMATION
As an additional aid in understanding APHM systems, the methodological processes involved in selecting the right AI techniques based on the type of input data are shown as a flow chart in Figure 3.As indicated in Figure 4 supervised methods, unsupervised, semi-supervised methods, transfer learning, and reinforcement learning are the five kinds of Artificial Intelligence techniques deployed for behavioral deduction of APHM system.Table 3 shows the various AI methods existing works based on factors.F I G U R E 4 Artificial intelligence learning methods for psychological health prediction.

| Supervised learning
The APHM system use supervised learning methods to determine an appropriate mapping between data and labels.
The role of a label in machine learning and deep learning is analogous to that of a dependent variable in traditional Using only accelerometer, we can achieve an accuracy of up to 69% for "userspecific models" and 61% for "similarusers" models.(Carneiro et al., 2012) SL Emotional states were classified using a J48 tree, a decision tree variant, which achieved an accuracy rate of 78% Lack of body language or gestures.
( According to this study, the subject's smartphone's accelerometer is utilized in conjunction with credulous Bayes models to determine whether or not a user's behavior is significantly correlated with their reported levels of stress (Garcia-Ceja et al., 2016).Emotional states were classified using a J48 tree, a decision tree variant, which achieved an accuracy rate of 78% in the study (Carneiro et al., 2012).Supervised methods such as convolutional neural networks (CNNs) and Recurrent Neural Networks (RNNs) are included in Deep Learning.
Automatic speech-based depression analysis.DCNNs learn deep-learning attributes from spectrograms and voice waveforms.Raw-DCNN uses audio signals and LLD, while Spectrogram-DCNN uses texture characteristics.They mechanized speech input and assessed depression severity using hand-crafted and deep-learned characteristics.Raw and spectrogram DCNN layer finetuning increased depression recognition.Shinde et al. (2020) propose a technique to diagnose depression using voice quality.Speech could be used to predict depression because voices and acoustics alter depression.After removing background noise, their method preprocesses 15 s audio samples.After audio segmentation, voice attributes are extracted using acoustic quality.Speech is spectrogrammed.CNNs need visuals.Their spectrograms interpret spoken motivations.They discovered a combination of acoustic elements that accurately classified a person's state (Dhyani & Kumar, 2021).To anticipate the chatbot's sentence into sentiments, M. Balaji et al. train the model with a preprocessed dataset and a bi-directional neural network (BRNN) with two hidden layers.As a preliminary step, BRNN (Balaji & Yuvaraj, 2019) makes long sentences from words.Chatbots learn to preprocess the content using Tokenization, data cleansing, and stop word removal, thus determining the sentence's sentiment, such as Positive, negative, and neutral based on modeling supervision.Model training provides basic chatbot labeled sentences.Python's Textblob module calculates statement sentiment from word sentiment scores.Happy, depressed, and neutral scores.User emotion determines the chatbot's PReLU tensor flow activation function's song.

| Unsupervised learning
In the context of the APHM framework, unsupervised learning frequently takes use of unlabeled data samples.In machine learning, the three types of unsupervised learning approaches that are utilized to identify data structures are dimensionality reduction, anomaly detection, and clustering.K-means and hierarchical clustering are two algorithms that accomplish this by grouping records into subsets with similar properties.One-class support vector machines and other anomaly detection techniques actively look for deviations from the norm.Attribute selection and principal component analysis are two examples of dimensionality reduction methods (Garcia-Ceja, Riegler, Nordgreen, et al., 2018) that can improve the generalization of machine learning models by eliminating multicollinearity and maintaining only the most important data.To prepare sensor data for analysis, personal sensing makes use of unsupervised learning techniques.The user's home or place of business, for example, can be located using a heat map generated from clusters of GPS locations.Using Bayes' theorem, we may create a classifier that takes into account the joint probability of data instances and labels and then compute the posterior probability (Garcia-Ceja, Riegler, Nordgreen, et al., 2018), it is possible to predict labels for fresh data instances.Scholarly studies such as (Garcia-Ceja et al., 2016;Xu et al., 2015) show how K-means clustering is used to identify stress in workplace conditions and groups of similar customers for the purpose of building improved stretch placement models.However, when it comes to predicting mental health with deep learning, autoencoders are by far the most widely used technique.Despite the abundance of touch data, text labels are in little supply.Specifically, we learn a low-dimensional representation of the heat maps that retains as much information as possible from the original heat maps, allowing us to make productive use of the data that has not been labeled (Diederik et al., 2015).This low-dimensional description (latent space or embedding) is extracted using a neural network design called a variational auto-encoder (Aksan et al., 2018).It autonomously construct meaningful low-dimensional representations in a wide range of domains.

| Semi-supervised learning
The technique of semi-supervised learning is utilized where there is just a limited quantity of named information available for prediction.Models are created using labeled and unlabeled occurrences in semi-supervised computations (Gopalakrishnan, Venkataraman, Gururajan, Zhou, & Zhu, 2022).Semi-supervised learning is crucial for Psychological state discovery/result since tagging input with the proper ground truth type is often challenging.Questionnaires are frequently used to predict daily disposition states; however, sometimes, participants do not answer, and a few days go untagged.To deal with bipolar clutter, labeling the information is done.A professional's opinion is frequently sought.This problem will limit the amount of tagged material retrieved (Maxhuni, Hernandez-Leal, et al., 2016).
Keyboard information (distance to the last keypress, time since the last keypress, and duration of a keypress), and accelerometer data have been used to make regression-based predictions of depression and mania in bipolar persons by other researchers (Huang et al., 2018).Computer keyboards now support dynamic keystroke attributes like pressure, latency, and duration.Epp et al. (2011) found that using these features, they could predict 15 states of emotions over two tiers with an accuracy of 77%-88%.Kołakowska (2013) summarizes the studies that have been conducted on the topic of determining a person's emotional state from their typing habits.

| Transfer learning
Transfer learning is a strategy for learning new information in a domain with little to no available labeled training data by drawing on experience gained in other, related domains where such data is available (Gu et al., 2023).For instance, one can use the skills honed while learning to recognize horses in photographs to identify cows.To be more specific, consider a computer software that first learns to recognize basic image categories like dogs, vehicles, horses, cats and then uses that information to learn how to determine whether or not an image has a specific disease.One of the drawbacks is that the findings could be better if the domains were more distinct from one another.In addition, it has yet to be thoroughly investigated in other scenarios outside of pictures and movies.Transfer learning has a significant potential for use in MHMS due to the fact that it has the ability to lessen the quantity of labeled data that is required, which is one of the primary obstacles that health applications face.Transfer learning, was utilized for detecting diabetic hypertensive retinopathy in Fundus images by Nagpal et al., when there was a scarcity of labeled data.Deep transfer learning, in conjunction with a neural network, was utilized by Alatrany et al. in order to perform early diagnosis of Alzheimer's disease utilizing scanned brain pictures (Alatrany et al., 2023).

| Reinforcement learning
A learning agent makes blunders and then fixes them as it interacts with its environment, a process known as reinforcement learning.The agent's objective is to achieve the most potential profit from the situation it is in by implementing the most beneficial course of action.In contrast to supervised learning, which requires the learning algorithm to be given the inputs and outputs that are to be targeted, reinforcement learning requires the agent to experiment with a variety of actions before selecting the one that results in the greatest accumulation of rewards.
When an agent takes the appropriate or inappropriate action, it is rewarded or punished in accordance with the situation.Over time, it makes an effort, to figure out which behavior leads to the best reward.One of the benefits of reinforcement learning is that it does not necessitate the participation of a human specialist who is knowledgeable about the application domain of the problem.In the past, applications of reinforcement learning have been utilized in various healthcare settings.For instance, the authors of Pineau et al. (2009) employed reinforcement learning to reduce both the frequency of epilepsy seizures and their duration.
Various algorithms are usually combined in practice to build effective final prediction models.Unsupervised learning algorithms, for example, are frequently employed as a preliminary step before constructing supervised learning models.In addition, a distinction is made between user-reliant and user-non-reliant ubiquitous training methods.The earlier are trained using information from the user in question.The model is tailored to each individual user based on their total user data, which is used in the process.User-dependent models can keep tabs on the unique qualities of each participant, allowing for more accurate results.Still, they necessitate many base classifiers for just that single user.Lu et al. (2012) studied both user-non-reliant and user-reliant models for the recognition of stress, and their results are a perfect example of this.

| Materials software packages/libraries
It is possible to analyze and train models with the help of a large range of specialized machine learning software tools and libraries.Others are supplementary libraries made for use with a particular programming language, while yet others are complete applications in and of themselves.The most popular machine learning tools are outlined in Table 4, which can be found here.Weka was utilized for the diagnosis of bipolar disorder in Alvarez-Lozano et al. (2014), Faurholt-Jepsen, Vinberg, Frost, Debel, Margrethe Christensen, Bardram, andKessing (2016), Grunerbl et al. (2014), Grünerbl et al. (2015), andDe Vos et al. (2016), for instance, in Zakariah and Alotaibi (2023) scikit-learn T A B L E 4 Software tools/libraries used for implementing the APHMS system.

Name of software tools/libraries Explanation
Matlab machine learning toolbox (Han et al., 2023;Wei et al., 2023) Matlab is a numerical environment that comes with various toolboxes, one of which is designed specifically for machine learning.MATLAB provides a variety of options for communicating with and transmitting data to various deep learning frameworks.
Keras (Kumar et al., 2023;Valtolina & Charlie, 2021) Google created the high-level application programming interface (API) known as Keras in order to facilitate the deep learning process.It is put to use in the process of implementing neural networks.It is a tool for implementing neural networks, and it was written in Python.The goal of this tool is to make the process of installing neural networks easier.In addition to this, it enables the computation of a wide variety of backend neural networks.
TensorFlow (Ly et al., 2017) TensorFlow is a library that may be used for a variety of different machine learning applications.
It is an open-source, end-to-end platform.
Scikit learn (Bendig et al., 2022;Zakariah & Alotaibi, 2023) Python library Scikit-Learn, often known as sklearn, is used to create machine learning models and statistical modeling.It is also known by its original name.We are able to create a variety of machine learning models through the use of scikit-learn, including models for classification, clustering and regression as well as statistical tools for assessing these models.
R (Gopalakrishnan, Venkataraman, Gururajan, Zhou, & Zhu, 2022) R is a programming language with an extensive library that includes libraries that implement many different types of machine learning strategies.R is a statistical programming language that is utilized for the purpose of data analysis and the graphical depiction of said data.R is well suited for studying statistics thanks to its extensive library of useful tools for data experimentation and investigation.
Spark Mlib (Abdullah et al., 2016) Spark's ML library is commonly known as MLlib.Its goal is to make true machine learning easier to implement at a larger scale.At its most basic, it provides resources like: Examples of ML Algorithms include the more common learning algorithms of collaborative filtering regression, clustering, and classification.This machine learning library scales well and works well with large datasets.
Weka (Alvarez-Lozano et al., 2014;Faurholt-Jepsen, Vinberg, Frost, Debel, Margrethe Christensen, Bardram, & Kessing, 2016;Grünerbl et al., 2015;Grunerbl et al., 2014;De Vos et al., 2016) This document contains a compilation of machine learning techniques that have proven useful in data mining.It is also possible to utilize it as an external library for Java projects, despite the fact that it features a graphical user interface (GUI).
was utilized for the recognition of anxiousness.However, authors rarely reveal the tools and libraries they employed.The next step to take after training a machine learning model is to make an estimate of how well its predictions will operate in actual real life situations, which is when the model is asked to make predictions based on data that it has not previously seen.

| Evaluation of artificial intelligence methods
In the APHM system, Artificial Intelligence methods are used in the evaluation to grade the execution in order to predict accurately.The preparing and testing sets are randomly assigned to one of the statics tests.The capacity to split data collecting into a prepare and test phase has been proven.Using k-overlay cross-approval, we create two subsets from the full dataset: one for training and another for validation.The progression described above is performed again for all subsets.Parameter tweaking is required for a few models.Data from the testing dataset could be introduced into the generated handling if these parameters are often adjusted depending on the execution of the validation set.When we have infinite amounts of data, delaying permission is fair.K-fold cross-approval is preferred when the total amount of information is limited.The machine learning show evaluation is used to grade the prepared show's execution to predict accurately.All of the other users' data is used in the training phase.In contrast, in psychological health wearable sensors, the most recent user's data is utilized as the test dataset to determine the algorithm's efficacy for new members.A usernon-resilient model, often called a generalist strategy, is one that does not require any particular set of inputs from the intended recipient.In particular, Miranda et al. employed a method that did not require any user input to detect signs of bipolar disorder in phone conversations (Mormann et al., 2005).In this instance, the same person owns the training and testing data.User-resilient strategies are then used to describe the model.Grünerbl et al. employed this assessment approach to identify depression and manic periods using smartphone motion traces (Hurling et al., 2007).
One strategy is to Common execution metrics for classification challenges including:

| Accuracy
A test's accuracy is determined by its ability to appropriately distinguish between healthy and sick instances.To calculate the fraction of genuine positive and true negative cases in all analyzed cases, to calculate the test's accuracy.Mathematically, this can be stated as: where FX is the amount of people who were incorrectly classified as patients; TY is the percentage of cases that were accurately classified as healthy; FY is the number of cases where healthy status was incorrectly assigned.

| Sensitivity
The ability of a test to appropriately identify patient instances is its sensitivity.To figure it out, we'll need to figure out what percentage of patient cases are true positive.Mathematically, this can be stated as:

| Specificity
Specificity measures how well a test can distinguish between unhealthy and healthy samples.Calculate the proportion of genuine negative in healthy cases to estimate it.Mathematically, this can be stated as:

| F1 score
It is indeed a combined recollection and precisely weighted average.For relapse challenges, the severe quadratic mistake, core cruel squared blunder, average relative error, association scores, and other implementation metrics are frequently used.It's crucial to note that using a single indicator to assess performance isn't a good idea (often only accuracy is used).Users can achieve a positive view of a classifier's or projection approach's genuine performance and resilience by using numerous measurements simultaneously.

| AUTONOMOUS PSYCHOLOGICAL HEALTH MONITORING (APHM) SYSTEMS' APPLICATIONS
All inter-sensor communications are managed by the middleware component of our Autonomous Psychological Health Monitoring (APHM) system, and the resulting data is made available to the applications in a standard data abstraction format as shown in Table 5.An application can only obtain sensor data by subscribing to the relevant broadcast channel.Electronic health records are responsible for communicating with sensor units, processing reading artifacts, and keeping tabs on the system's vitality.Three case studies are based on mental health monitoring, each drawing inspiration from a different field of use: care for the elderly, fitness support, and tele monitoring.

| Remote monitoring
The application designed to aid the remote monitoring keeps tabs on the user's vitals and interactions with the surroundings, which are interpreted as ADLs.The system sends a notification if the user fails to complete an action (like brushing their teeth) within a predetermined time window.It also keeps track of users' daily activities to see whether your routine has shifted.The aid application employs low-power Wi-Fi modules outfitted with a ballin-tube sensor, a reed switch, or a passive infrared sensor in addition to body sensor networks.The modules are described in detail in Boopathi (2023).When motion is detected, the sensor units communicate this information to the phone via HTTP POST messages, which are then processed by rules to infer the appropriate responses.When the cutlery drawer and the refrigerator are opened within 45 min of each other, the app will infer that someone is eating if the sensor module-equipped dining room chair is occupied.Similarly, mental health monitoring systems keep track of things like when a person brushes their teeth, showers, opens windows, eats, cooks, and works at a desk.The technique does not damage the phone's Wi-Fi capabilities because it uses of the current Wi-Fi network.

| Well being tracking
Physical activity can be considerably increased and maintained using an online and phone-based user motivator system, according to studies (Hurling et al., 2007).Like physical health monitoring systems may be used to develop a wellness app, mental health monitoring systems can do the same.That keeps tabs on the user's heart rate and records their daily activities (Prochaska et al., 2021).Two separate sensor network configurations are included: one for tracking regular activities and another for recording workouts.The user wears a single accelerometer, and three independent states are identified: resting (idle) (sitting, standing), light activity (walking, cycling), and vigorous activity (running, jumping, etc.).(i.e., running).Different weightlifting gloves and chest strap sensors enable activity classification and rep counting for 16 other routines.An analysis of the system indicated that the application was capable of accurately keeping track of activity and rhythm of the heart in real time with input from the user for a minimum of 12 h each day (Seeger et al., 2011), and its recognition performance was on par with state-of-the-art methods.

| Emergency/tele medicine
The field of telemedicine and emergency notification is one promising area for the application of body sensor networks because it allows for the remote monitoring of patients for extended periods of time and the remote detection of emergencies, such as falls.A number of sensors monitor the user's weight, heart rate, as well as their everyday movements, electrocardiogram (ECG), and cholesterol levels.This program also tracks the user's electrocardiogram (ECG).The user's heart rate can be linked to the motions in order to trigger an alarm if the user's heart rate goes outside of the typical range for that activity (Sharifi Kia et al., 2023).Information gathered by sensors is uploaded to a database and later shared with a preexisting telemedical system.User-initiated measurements (like those made by a Bluetooth-enabled scale) trigger notifications (Zhong et al., 2023).

| Emotion monitoring
Emotions are mental states that can't always be detected by looking at the same exterior indicators.Daily ecological momentary assessments (EMAs) found a strong correlation between students' use of built-in sensors for mental health monitoring like Bluetooth-detected contacts, SMS messaging, calls and location entropy (an assess of the spatial dispersion of locations).A sensor system that is based on a smartphone might theoretically record paralinguistic characteristics such as the speaker's amplitude and conversational cadence, and then use this data to assess the user's amount of stress that they are currently experiencing (Calvo & D'Mello, 2010).Computer mouse and keyboard interactions may be used to detect emotions like stress based on attributes that can distinguish between a laboratory-induced stressful state and a normal condition.Whether or not these interactions generate a sufficiently strong stress signal in real-world examples remains an unanswered subject (Ciman et al., 2015).Support, psychoeducation, psychological treatment, tracking behavioral adjustments, and relapse prevention are just some of the many uses for chatbots in digital mental health interventions (DMHIs) (Alattas et al., n.d.).

| Challenges
Though there are substantial advances in Autonomous Psychological Health Monitoring (APHM) systems, there are still challenges to solve before fully automated and appropriate systems can be adopted.These challenges bring several investigation ways to enhance aspects of the many stages of creating autonomous psychological state surveillance systems.Table 6 illustrates the summary of challenges in emotional well-being systems.Following that, some research problems and potential exist in the psychological health monitoring system.

| Labeling data
Tagging data or aligning sensor information with the relevant factual psychological state is one of the most challenging tasks in developing APHM systems.Learning algorithms are constructed using this "ground truth" information.The accuracy of the model's projections is proportional to the precision of the input data.In order to detect and characterize stress, Maxhuni, Hernandez-Leal, et al. (2016) employed transfer learning due to a lack of manually annotated data.Another problem with self-reported judgments is the uncertainty of the data.Participants may be expected to describe their situation through a survey for off-site analyses.It depends on consciousness, is subject to prejudice, and seizes mistakes due to ego.Such inaccuracies will have a significant influence on estimation techniques.Although there are methods for managing labeled confusion (Frénay & Verleysen, 2014), they have scarcely been used in tracking cognitive processes.

| Integration of various sensors data
Data in APHM systems can be gleaned from information such as the system's software and hardware versions.Due to differences in presentation, measurement units, sample frequency, and so on, preparing data from different sources requires multiple stages.Data used in analytics for predicting needs to be in a standard format.One common strategy for doing so is called "gathering" or "early integration," which involves creating vectors that include information from all available senses.The classifier is trained using these combined vectors, which are then used for inference.The training process can be slowed if too many characteristics are extracted, which is a potential drawback of this approach.
When the classifier is trained on the server side, this is not an issue, but when it is done locally on a smartwatch, it can take a long time and drain the battery.Because each sensing method would have its attribute values and the number of instances, sensors with far fewer characteristics would be overlooked in the accumulation strategy.Exclusively building models for each sensor and then merging their results to achieve the ultimate class is one example of a delayed fusion technique.Soleymani et al. (2017) employed late integration to integrate multiple sensor types (pupil dilation, body language, user engagement with the system, oxygen consumption, galvanic numbing sensation, and visual attributes to obtain superior results in their studies on picture search intent recognition.

| User variance
Two types of variation can exist in APHM systems: inter-user variance and Intra user variance.Physiological and behavioral trends will differ amongst users due to the individual nature of each person.As a result, user-non-resilient models outperform user-dependent models.Muaremi et al. found that the user-non-resilient and user-resilient models had 53% and 61% accuracy, respectively, in their stress diagnosis study (Garcia-Ceja, Riegler, Nordgreen, et al., 2018).Intra-user fluctuation can be caused by genetic factors such as metabolic activities, speech changes during menopause, and illnesses.Research on women's stress and language-based emotion identification found that intra-user variation fluctuates with the menstrual cycle (Giddens et al., 2013).The participants' habits are changing in a short duration.The authors in Servia-Rodríguez et al. ( 2017) discovered multiple spatial unsociable hours on weekdays and weekends using the GPS tracking system, cell towers, and wireless broadband data.Due to intra-user instability, a user-dependent model currently performing well can begin to deteriorate over the period.Dynamic time approaches are ideal in these situations.Conceptual transition is an issue that occurs when the correlation between source data and variable output modifications, and in the domain of supervised learning, numerous adaptive techniques have been presented (Gama et al., 2014).As far as we understand, neither preceding Psychological well-being observation research has explored the intra-user variance problem.

| System interoperability
While most of our discussion so far has focused on machine learning-related challenges, remember that these APHM devices are designed to be compatible with a wide range of apps and, eventually an Internet of Things environment.Web-based (Ybarra & Eaton, 2005), mobile phone-based (Diraco et al., 2023), and other healthcare intervention and support systems include databases for caregivers and physicians, system interfaces, and administrative tools.If services are to be reliable and scalable, channels between modules and platforms must be acknowledged and developed.

| Inadequacies of artificial intelligence-based message-exchanging robots
In the field of APHM systems, chatbots have been shown to have a number of drawbacks, which should not be confused with a Utopian solution to the issues that now exist with digital mental health interventions (DMHIs).For instance, chatbots often misinterpret user responses because of the ambiguity inherent in human language (Adamopoulou & Moussiades, 2020).As more languages are added to the mix, it will become increasingly difficult for chatbots to understand figurative language such as metaphors, ellipses, and exaggeration (Neidlein et al., 2020).One of the most consistent criticisms of chatbots heard across multiple research is that consumers find their interactions with them to be boring after a while (Bickmore et al., 2005;Fitzpatrick et al., 2017;Fulmer et al., 2018;Ly et al., 2017).The chatbot's lack of humanity is one reason why people stop using it after a while (Adamopoulou & Moussiades, 2020).Concerns about misunderstandings have been raised by users (Fitzpatrick et al., 2017;Inkster et al., 2018), especially when longer or more complicated interactions are involved.This can lead to inappropriate responses that damage the therapeutic relationship (Kretzschmar et al., 2019).Chatbots, despite the fact that they may have some useful applications, may discourage consumers from interacting with the intervention in a number of different ways.When chatbots or other types of artificial intelligence are employed for medical diagnosis, potential hazards of misconceptions that are already present are magnified.

| Potential opportunities
Autonomous Psychological Health Monitoring (APHM) systems have come a long way, but they still face several obstacles that present chances for study.When designing a system to track a person's Psychological health, data labeling is an essential first step for determining whether or not the collected sensor data accurately reflect the person's actual Psychological state at a given time.Consequently, it is crucial to investigate machine learning algorithms further to assess their potential for Psychological health monitoring.Inter user variations are another vital consideration for physiological and behavioral tracking.Hybrid models can include the benefits of user-dependent and user-independent (generic) approaches, as previously indicated (Xu et al., 2015;Zenonos et al., 2016).So, it is important to investigate the possibilities of this hybrid paradigm.Also, Intra user variances, which relate to variations in a single user's physiological and behavioral characteristics, must be considered.At the end of the day, the most effective means of communication will be those backed by clinical evidence and integrated with patient databases, managerial resources for caregivers and physicians, and existing support systems.Monitoring Psychological health requires using multimodal sensing technology and suitable machine learning approaches.Few studies, however, have used a control group or compared the effects of engaging with a chatbot to those of doing something else, like reading a book with psychoeducational content.It is unclear how much chatbots improve the efficacy of DMHIs because control groups in the vast majority of trials, if not all of them, consisted of insufficient mimic circumstances.Keep in mind that this information will be useful for clearing up misunderstandings about the scope and limitations of AI chatbots in this setting.Researchers and developers working in the field of AI with the goal of using it to digital medicines need to improve their ability to work together and share information.How much artificial intelligence (AI), digital therapeutics sector utilizes in their interventions will define how patient-focused their chatbot creation and referral processes are.

| Limitations of this survey
The primary objective of this study was to conduct a comprehensive literature review with the goal of outlining the current state-of-the-art in Autonomous Psychological Health Monitoring (APHM) systems.These systems make use of multimodal sensing methods, and virtual agents to collect and process the physiological information from the raw data, and various artificial intelligence methods such as machine learning and deep learning algorithms, amongst others used to deduce the data into the meaningful emotional states predictions.The review was intended to provide an overview of the current state-of-the-art in APHM systems.
• This work does not provide a comprehensive assessment of each case but selects relevant works for many situations in the context of Autonomous Psychological Health Monitoring (APHM) systems hardware and software.• Section 2 provided a high-level summary of APHM's most often employed phases; nevertheless, certain works may have utilized more or fewer processes.For the sake of simplicity and breadth, we endeavored to collect the most common ones.

| Future issues
1.It is anticipated that a sizable user base will be necessary for widespread adoption of personal sensing systems.This is due to the wide range of outcomes that may arise from technology, device-use patterns, lifestyle, and environmental factors.2. Information gathered by GPS trackers, for example, can never be de-identified without completely destroying its utility.Therefore, it is important to provide users with tools to understand, control, and take credit for their data if we want them to feel safe using these systems.3. The development of interventional technologies based on personal sensing applications is possible.Consumers expend less effort since these technologies produce contextualized and tailored interactions for them.4. Researchers, developers, and users need to agree on how much error can be allowed and the best approach to characterize and present error to key stakeholders because no sensing system can be guaranteed to have a precision of one hundred percent. 5. Users will need to put in some work to help improve systems.More interesting and empowering systems are likely to be developed if methods are discovered to assure that behaviors are directly associated with rewards.6.The integration of personal sensing into the workflow and the development of supporting infrastructure are necessary for its potential to enhance disease detection and treatment access.There is also a need for more algorithmic precision, in addition to developments in underlying technology and understanding.7.There will always be a gap between ideal and realistic goals in the realm of personal sensing.There is friction between small studies that prove the concept and show novelty, on the one hand, and large studies that show resilience and generalizability, on the other.8. Integrating personal sensing data with clinical and genetic databases can help shed light on the role of behavior and gene-behavior connections in health, wellbeing, and disease.

| CONCLUDING REMARKS
Recent state-of-the-art research on Autonomous Psychological Health Monitoring (APHM) systems is reviewed in this article.Most emphasis is on studies that use sensors and virtual chatbots to collect behavioral data and artificial intelligence methods to interpret these data into higher level attributes to predict Psychological diseases.Also, we recognized the primary stages of APHM systems, beginning with the design of the experiment and ending with the implementation of the system (Section 2).They include a significant taxonomy of the factors and concerns for the observation of data gathering and processing, and the training and assessment of learning models.There is growing evidence that machine learning can use data collected from a smartphone's sensors to identify indicators of rest (such as bedtime or waketime, duration), emotions, social factors (such as who is around and their relationship to contacts in the smartphone), and stress.Along with that, various applications deployed using APHM systems are discussed.We addressed some of the difficulties that have been encountered in APHM systems research as well as some of the potential that lies ahead for the area.According to the reviewed literature, integrating multimodal sensing technologies with artificial intelligent learning methods gives a significant possibility for further development of autonomous monitoring Psychological health care technology tools for treatment.Researchers and producers of monitoring systems are interested in what factors encourage regular use, continuous trust, and persistent use from end users.Though individualized perceiving is still in its early stages, it has enormous potential as growth of mental health issues, way for undertaking research on mental health as well as a diagnostic instrument to track populations at risk and establishing the groundwork for future generations of autonomous monitoring health solutions.

F
I G U R E 3 Flowchart of AI methodologies in autonomous psychological health monitoring (APHM) systems.
Publication citations are indicated in the column headers.The columns represent the many aspects and questions asked in each survey.Applications: EM, emotion monitoring; Emg, emergency; RE, remote monitoring; WBT, well being tracking.
Relationship between data source, behaviors, and attributes.
T A B L E 1Note: X is not applicable.

Observation Factor's Taxonomy Observation Study Observation Duration Observation Device Correlation Prediction Detection Short Term Long Term Smart phone Ambient sensors Wearables Virtual agent chat Observation Phenomena Anxiety disorder Depression Bipolar defective disorder
Summary of the reviewed papers under the observation categories.

Artificial Intelligence learning methods
T A B L E 3 AI methods based on factors.
modeling.Personal sensing relies on user-reported labels.Data instances and their respective labels make up training samples.Labels act as a tutor who keeps an eye on training data.The trained mapping function can then draw inferences about the labels from the unlabeled data.Classification and regression are two supervised learning examples in machine learning.The values used by regression methods and functions are often continuous.In supervised learning, classification is typically used.Sensor readings are classified into activities like walking, running, and sleeping via activity recognition.Classification algorithms can be either generative or discriminative.
Note: In type of data: RL, reinforcement learning; SL, supervised learning; SSL, semi supervised learning; TR, transfer learning; USL, unsupervised learning.statistical Summary challenges in psychological health monitoring.