Machine learning based on Optical Coherence Tomography images as a diagnostic tool for Alzheimer's disease

Abstract Aims We mainly evaluate retinal alterations in Alzheimer's disease (AD) patients, investigate the associations between retinal changes with AD biomarkers, and explore an optimal machine learning (ML) model for AD diagnosis based on retinal thickness. Methods A total of 159 AD patients and 299 healthy controls were enrolled. The retinal parameters of each participant were measured using optical coherence tomography (OCT). Additionally, cognitive impairment severity, brain atrophy, and cerebrospinal fluid (CSF) biomarkers were measured in AD patients. Results AD patients demonstrated a significant decrease in the average, superior, and inferior quadrant peripapillary retinal nerve fiber layer, macular retinal nerve fiber layer, ganglion cell layer (GCL), inner plexiform layer (IPL) thicknesses, as well as total macular volume (TMV) (all p < 0.05). Moreover, TMV was positively associated with Mini‐Mental State Examination and Montreal Cognitive Assessment scores, IPL thickness was correlated negatively with the medial temporal lobe atrophy score, and the GCL thickness was positively correlated with CSF Aβ42/Aβ40 and negatively associated with p‐tau level. Based on the significantly decreased OCT variables between both groups, the XGBoost algorithm exhibited the best diagnostic performance for AD, whose four references, including accuracy, area under the curve, f1 score, and recall, ranged from 0.69 to 0.74. Moreover, the macular retinal thickness exhibited an absolute superiority for AD diagnosis compared with other enrolled variables in all ML models. Conclusion We identified the retinal alterations in AD patients and found that macular thickness and volume were associated with AD severity and biomarkers. Furthermore, we confirmed that OCT combined with ML could serve as a potential diagnostic tool for AD.


| INTRODUC TI ON
Alzheimer's disease (AD) is the most common neurodegenerative disease and is the leading cause of dementia. It is pathologically characterized by extraneuronal amyloidβ (Aβ) plaque deposition and intracellular neurofibrillary tangles consisting of hyperphosphorylated microtubule-associated protein tau or tau. 1 It has been reported that the prevalence of AD continues to rise worldwide, posing a serious burden on societies and their families. 2 Studies have suggested that changes in positron emission tomography (PET) and cerebrospinal fluid (CSF) biomarkers can be detected years before the emergence of symptoms. 3 Currently, the diagnosis of AD is reliant on detecting decreased Aβ 42 and elevated tau levels in the CSF, 4 cerebral glucose hypometabolism, and increased Aβ and tau deposition on PET images. 5 However, these modalities are invasive, time-consuming, and expensive; therefore, impractical for mass screening for AD. Thus, to achieve early diagnosis and intervention, it is necessary to develop a rapid, noninvasive, and straightforward tool to identify AD.
The retina might be an important site of inquiry because it shares the same embryonic precursor as the brain and displays several structural and functional similarities with the organ. 6 Previous research suggests that visual symptoms, including impaired visual acuity, visual field, motion perception, and stereopsis, are frequently early symptoms of AD. 7 In the past, these symptoms have been attributed to degeneration of the central visual pathways. However, recent studies have shown that the retina is also involved. 8,9 Additionally, Aβ deposits have been found in the retina of AD patients and mouse model. 10 High-definition optical coherence tomography (HD-OCT) is a costeffective and convenient tool that allows for high-resolution retinal images to be obtained in vivo. With the advent of OCT, the retina has become an appealing subject in many studies on neurodegenerative diseases, including AD. [11][12][13][14][15][16][17] However, previous OCT findings are controversial; some studies have reported a thinning of the retina in AD, [11][12][13][14] whereas others have observed no change. [15][16][17] While there has been much interest in exploring the potential correlation between dementia and retinal parameters, there is a paucity of comprehensive and consistent data from currently available studies.
Artificial intelligence (AI), particularly machine learning, has recently emerged as a powerful tool in the field of neuroscience. It refers to the ability of computers to detect complex patterns and construct models by learning from existing data. Machine learning has been developed and applied extensively in the field of medicine, including disease diagnosis and prognosis, disease progression monitoring, and treatment efficacy evaluation. [18][19][20] Therefore, in this study, we first evaluated retinal alterations in patients with AD and then assessed the relationships between the retinal structure and the clinical parameters, as well as AD biomarkers. Finally, we evaluated the performance of six common machine learning algorithms based on OCT images.  both groups, the XGBoost algorithm exhibited the best diagnostic performance for AD, whose four references, including accuracy, area under the curve, f1 score, and recall, ranged from 0.69 to 0.74. Moreover, the macular retinal thickness exhibited an absolute superiority for AD diagnosis compared with other enrolled variables in all ML models.

Conclusion:
We identified the retinal alterations in AD patients and found that macular thickness and volume were associated with AD severity and biomarkers.
Furthermore, we confirmed that OCT combined with ML could serve as a potential diagnostic tool for AD.

K E Y W O R D S
Alzheimer's disease, diagnosis, machine learning, optical coherence tomography, retina All procedures followed the Declaration of Helsinki and were approved by the ethics committee of Xiangya Hospital (approval no. 201811199).

| HD-OCT examination
All OCT scans (Cirrus HD-OCT 4000; Carl Zeiss Meditec, Dublin, California) were performed by the same ophthalmologist, who was blinded to the diagnoses according to standard protocols. Due to time constraints and a high inter-eye correlation between both eyes, we selected one random eye for the HD-OCT examination.
The thickness of the peripapillary retinal nerve fiber layer (pRNFL) was acquired by taking three consecutive 3.4-mm circular scans centered on the optic nerve disk (Figure 2A). The macular parameters were measured by taking six consecutive 6-mm radial line scans, each containing 128 A-scans within a 6 × 6 mm macular region.
The macular retinal thickness (MRT) was measured in nine regions corresponding to the procedures of the Early Treatment Diabetic Retinopathy Study (ETDRS) ( Figure 2B). The pRNFL thickness (mean, superior, inferior, nasal, and temporal), MRT averaged over all nine ETDRS subfields, and total macular volume (TMV) were analyzed in this study. We excluded participants with HD-OCT images of poor quality (signal strength <7) and those who could not follow OCT procedures due to severe cognitive impairment.

| Macular intraretinal layer segmentation
All OCT images were segmented using advanced automated three-dimensional retinal layer segmentation software (IOWA OCTExplorer v3.8.0). 22 Segmentation was performed by two independent computer engineers who were blinded to the diagnosis of the participants. Briefly, this algorithm outlined 11 optical surfaces, from the internal limiting membrane to the retinal pigment epithelium, and automatically segmented the retina into the following 10 layers: retinal nerve fiber layer (mRNFL), ganglion cell layer (GCL), inner plexiform layer (IPL), inner nuclear layer, outer plexiform layer, outer nuclear layer, inner segments/outer segments junction, outer segment, outer segment photoreceptor/retinal pigment epithelium junction, and retinal pigment epithelium layer ( Figure 2C).
The thickness of the intraretinal layer was calculated for each of the nine ETDRS subfields. The average macular intraretinal layer thickness of the nine ETDRS grids was then calculated for further analysis.

| Magnetic resonance imaging and CSF biomarkers
All patients underwent 3-Tesla MRI scans on the same scanner (Magnetom Verio; Siemens, Erlangen, Germany). The severity of F I G U R E 1 Flow chart of participants recruitment into the study. A total of 159 patients with AD and 299 age-and sexmatched HCs were included in the final analysis. AD, Alzheimer's disease; HCs, healthy controls. brain atrophy was assessed using three rating scales: Medial temporal lobe atrophy (MTA), 23 global cortical atrophy (GCA), 24 and Koedam's scale for parietal cortical atrophy (PCA). 25 MR images were scored by an experienced scorer who was blinded to clinical information.
Of the 159 patients with AD, 48 consented to lumbar puncture for CSF biomarker testing. CSF samples (5-15 ml) were centrifuged at 2000 × g for 10 min and stored at −80°C, according to a previously established protocol. The levels of Aβ 42 , phosphorylated tau (p-tau), and total tau (t-tau), as well as the Aβ 42 /Aβ 40 ratios in the CSF, were measured using enzyme-linked immunosorbent assays. 26

| APOE genotyping
Of the 159 patients with AD, 10 refused to provide their blood samples. Thus, we assessed APOE genotyping in 149 patients with AD. APOE genotyping was performed on DNA extracted from a 10 ml blood sample using previously established protocols. 27 The measurements were performed in a blinded manner.

| Statistical analysis
Data are expressed as mean ± standard deviation, median or percentages. The Kolmogorov-Smirnov test was used to determine the normality of the data. Student's t-tests and Chi-square tests were used to assess the difference between two variable groups when the sample data were normally distributed; otherwise, the non-parametric test was used. The covariance analysis was used to compare OCT measurements between AD patients and HCs, with age, sex, IOP, BCVA, and axial length (AL) taken as covariates. Similar analyses were performed for comparisons across AD subgroups and each subgroup versus HCs. We evaluated the correlation of OCT parameters with cognitive function, brain atrophy, CSF biomarkers, and APOE genotypes by using Pearson correlation analysis.
Statistical significance was set at p < 0.05. All statistical analyses were performed using SPSS Statistics, version 25.0 (IBM Corp.).

| The performance evaluation of different models
OCT variables with a significant p-value at the above statistical analyses were incorporated into the final diagnosis model. We We randomly selected 70% of the dataset to obtain the models and used the remaining 30% of the dataset to test the performance of the models. Four criteria were used to evaluate the performance of each model, including the accuracy (ACC), the area under the curve (AUC), f1 score, and recall. We used fivefold cross-validation to validate the stability of these constructed models, and the accumulated ACC, AUC, f1-score, and recall values were used as the metrics for evaluation. All the construction processes of prediction models were based on Python programming language (version 3.7).

| Characteristics of study participants
We included 159 patients with AD and 299 HCs in the final analysis. We found no statistically significant differences in age, sex, or ophthalmologic parameters (IOP, BCVA, and AL). The AD group had a shorter mean duration of education (p < 0.001) and lower mini-mental state examination (MMSE) scores (p < 0.001) than the HCs group (Table 1).

| HD-OCT measures in AD and healthy controls
After adjusting for potential confounders, the average, superior, and inferior quadrant pRNFL ( Figure 3A), mRNFL, GCL, and IPL ( Figure 3C

| Correlation between OCT measures and cognitive performance
We

| Correlations between OCT measures, MRI, and CSF biomarkers
Correlations between OCT measures and brain atrophy assessed by MTA, PCA, and GCA scores were also analyzed. After adjusting for potential confounders, we observed a significant inverse associa-

| DISCUSS ION
This study revealed that OCT measurements were significantly cor- and their fibers are located in the macular area, a region in which progressive loss and degeneration of RGCs have been detected in AD patients. 31 The retinal Aβ deposits are more concentrated in the inner macular layer (mRNFL, GCL, and IPL), leading to the thinning of these layers. 31 It could also be due to the degeneration of RGCs and their axons or the retrograde trans-synaptic degeneration of the RGC layer and its axons. 32 Furthermore, the degeneration of RGC dendrites confined to the retinal IPL preceded cell loss in an AD mouse model, 33 indicating that the inner retinal layer may prove to be a potential site of interest for early AD detection.
We found that retinal measures were positively associated with MMSE and MoCA scores. This suggests that OCT parameters correlate with cognitive function in patients with AD. Oktem et al. 34 also found there was a significant correlation between RNFL thickness and MMSE scores. In addition, Iseri et al. 35  retrograde retinal atrophy may occur in AD. 37 The cingulate cortex F I G U R E 6 The weight plots of different variables from five machine learning models. The length of the bar indicates the importance of the variable. To facilitate a more intuitive comparison, this most important variable was taken as the baseline reference value, and other variables were presented as percentages of the reference value. There is no weight plot from KNN, which is a clustering algorithm, and thus is not amenable to linear fitting.
is the central hub of cognitive brain networks, particularly the posterior cingulate cortex, which is important for processing episodic memory tasks. 38 The findings from recent studies further suggest that the hippocampus and cingulate cortex play key roles in the neurodegenerative processes emblematic of AD. However, we found no significant associations between OCT measures and GCA scores. Again, this is likely because of the high proportion of mildto-moderate AD patients in our dataset (74.21%). The global brain structures of such patients are often preserved; therefore, few patients with AD might have developed whole-brain alterations in our study.
Our study showed that GCL thickness was positively correlated with CSF Aβ 42 / Aβ 40 and negatively correlated with CSF p-tau levels.
Additionally, IPL thickness was positively correlated with CSF Aβ 42 levels and negatively correlated with CSF t-tau levels. There are few reports on the association between retinal thickness and CSF biomarkers. A study found that neither pRNFL nor macular thickness was associated with CSF biomarkers. 16 However, their study sample size was small, including only 15 patients with AD. In postmortem studies, Aβ deposits and p-tau protein have been observed in patients with AD. 39 We hypothesize that the levels of Aβ and tau in the retina mirror those in the brain. A lower CSF Aβ 42 and a higher tau concentration reflect a higher Aβ and tau burden in the retina and brain. Thus, the retinal pathophysiological changes could eventually lead to decreased retinal thickness.
We found that all retinal parameters were decreased in the APOE ε4-carrier group compared to the APOE ε4-noncarrier group, although this difference was only significant in terms of IPL thickness.
In previous studies, patients with AD carrying APOE ε4 have been shown to exhibit increased apoptosis and loss of synaptic integrity. 40 In a mouse model, APOE ε4 was also associated with reduced microvascular density and neocortical cerebral blood flow. 41 Retinal capillary density is associated with brain perfusion and is reduced in early AD. 42  Regression. This confirmed that machine learning had a wide range of application prospects in clinical practice. Because of its ability to detect complex patterns and construct models, it was available to assist in the clinical diagnosis of AD, and thus, achieve the aim of the early invention and optimizing treatment strategies. Further studies are required to validate these findings.
This study had some limitations. Since OCT measurements and neuropsychological assessments require patients' cooperation, patients with more advanced stage AD (CDR = 3) could not be included in our study. Additionally, this was a cross-sectional study; therefore, we cannot conclude retinal changes over time. However, we compared OCT measurements among patient subgroups according to disease severity and HCs groups, which roughly simulates the process of retinal changes over time. Moreover, MRI and CSF analyses were not available for the HC group.
In conclusion, this large-scale study confirmed that retinal structure was significantly altered in patients with AD and that OCT measurements correlated with cognitive function, MRI findings indicative of cerebral atrophy, tau and Aβ levels in the CSF, and APOE genotypes. In addition, we developed an optimal machine learning algorithm to assist the clinical diagnosis of AD. Overall, this study confirmed that OCT measurements combined with machine learning could be useful for AD diagnosis.

ACK N OWLED G M ENTS
We are grateful to all participants in the present study.

CO N FLI C T O F I NTE R E S T
The authors have no competing interests to declare.

DATA AVA I L A B I L I T Y S TAT E M E N T
All data that support the findings of the current study are available from the corresponding authors upon reasonable request.