Risk of Pseudomonas aeruginosa antimicrobial resistance using time series analysis of antibiotic usage

Antimicrobial resistance is emerging as one of the most potentially disastrous threats of the 21st century. Pseudomonas aeruginosa (P. aeruginosa) is a leading resistant pathogen and is clinically significant due to its limited available treatment using antibiotics. Rising resistance of P. aeruginosa is of increasing concern and it is currently listed as one of the top three critically resistant pathogens by the World Health Organization. It is currently known that resistance in P. aeruginosa is significantly linked with the consumption of all antibiotics, making usage surveillance of particular concern.


INTRODUCTION
Antimicrobial resistance (AMR) is a significant health concern and occurs when microorganisms develop the ability to withstand the drugs designed to kill them through resistance mechanisms.This leads to prolonged hospitalisation and overburdened health systems and is predicted to cause 10 million deaths annually by 2050. 1 In 2019, there was an estimated 4.95 million deaths associated with bacterial antibiotic resistance. 1 The six leading pathogens, including Pseudomonas aeruginosa (P.aeruginosa), were responsible for 3.57 million of these deaths, and P. aeruginosa alone was responsible for more than 50 000 deaths. 1 P. aeruginosa is a gram-negative bacterial pathogen that commonly presents in hospitalised or immunocompromised patients.It can cause urinary tract infections (UTIs), septicaemia, chronic persistent airway infections, and is associated with burn and wound infections. 2 The increase in multi-drug-resistant strains of P. aeruginosa limits treatment options, resulting in inappropriate empiric therapy with poor clinical outcomes. 3Due to this burden, the World Health Organization (WHO) lists carbapenem-resistant P. aeruginosa as one of three bacterial species of critical priority for new antibiotic development. 3Importantly, resistance has been observed in P. aeruginosa against all currently available antipseudomonal antibiotics due to strong intrinsic, acquired, and adaptive resistance. 4. aeruginosa possesses a range of intrinsic resistance mechanisms, including a semi-permeable outer membrane, efflux pumps, degrading enzymes, and drug modification.These intrinsic factors are of clinical significance due to their ability to restrict entry of b-lactamantibiotics, transport various antibiotics out of the cell from the periplasm, enzymatically degrade antibiotics through hydrolysis, and modify drug targets to inactive forms. 3cquired mutation in P. aeruginosa involves horizontal transfer via conjugation, transformation, or transduction.This is clinically significant as it allows the transfer of resistance mechanisms. 2 The major adaptive mechanism is exposure to subinhibitory concentrations of antibiotics.This causes over-expression of genes encoding efflux pumps, increasing resistance.
Studies show a direct correlation between antibiotic usage (AU) and AMR, and this association is well accepted.Large-scale community AU is especially linked to increased resistance observed in hospital settings, suggesting AU is a driving factor in AMR. 6AU surveillance is thus essential in monitoring resistance in P. aeruginosa.[9][10] To model and project trends in AU, time-series forecasting was used.Time-series forecasting is the process of predicting future trends based on known past events, involving transforming a time series into a form that standard machine learning can process. 11This was used to forecast future AU trends from existing antipseudomonal prescription data and evaluate the impact of AURA using interrupted time-series analysis.
The aim of this study was to model trends, using time-series analysis, of current and future antibiotic usage through available prescription data for antipseudomonal antibiotics.

METHOD Ethics Statement
Ethics approval was not required for this research article as it used publicly available data and did not include human subjects.

Overview
A retrospective descriptive analysis on the prescription data of gentamicin, tobramycin, cefepime, and ciprofloxacin in Australia was undertaken using time-series methods.These data ranged from 1 January 2000 to 1 June 2020.

Data Collection
Antibiotic usage was collected through monthly prescription records of gentamicin, tobramycin, ciprofloxacin, and cefepime from the Pharmaceutical Benefits Schedule Item Reports of Australian Government Services. 12The item numbers for each drug dosage were found under the Pharmaceutical Benefits Scheme (PBS). 13The obtained prescription data were typed as either PBS or RPBS (Repatriation Pharmaceutical Benefits Scheme).Of the four antibiotics, there were 21 different dosages: 6 for ciprofloxacin, 2 for cefepime, 10 for tobramycin, and 3 for gentamicin.The PBS, RPBS, and total services of each of these dosages were transformed into 21 separate Microsoft Excel (Microsoft Corporation, Redmond, WA, USA) spreadsheets.
Of the 21 dosage forms available, 12 forms were able to be transformed into defined daily doses (DDDs) and were investigated further including 3 for ciprofloxacin, 2 for cefepime, 6 for tobramycin, and 1 for gentamicin.Therefore, the final prescription data used for analysis were retrieved from items 1210Q, 1209P, 1208N, 8316Q, 8315P, 9480Y, 8872Y, 5442K, 1356J, 10074F, 10066T, and 2824P.To calculate a comparable unit between dosages, services were divided by the appropriate DDD retrieved from the Anatomical Therapeutic Chemical (ATC)/DDD index. 14The unit, DDD/1000p, was then calculated using annual populations from the Australian Bureau of Statistics. 15The monthly DDD/1000p of each dosage were added to yield a separate spreadsheet for the total DDD/1000p for tobramycin, gentamicin, cefepime, and ciprofloxacin, and then for a final total antipseudomonal DDD/1000p sheet.

Data Analysis
Facebook Prophet time-series methods on Python 3.9.14(Python Software Foundation, Wilmington, DE, USA) was used to analyse and forecast trends to 2025.This was modelled based on the prior trends in the monthly DDD/1000p.Only PBS usage was analysed due to inconsistent and scarce RPBS data that had insignificant contributions.
Prophet involves open-source time-series forecasting and outperforms other modelling methods, such as autoregressive integrated moving average (ARIMA), when dealing with multiple seasonal effects. 16It is based on a univariate generalised additive model (GAM) and involves a curve-fitting task with time as the regressor.The model contains a trend, seasonality and an external variation component.Seasonality is accounted for using Fourier series.Prophet fits its GAM using the L-BFGS quasi-Newton optimisation method, automatically producing monthly predictions of DDD/1000p in its forecast. 17sing prescription data from January 2000-June 2020 as training data, Prophet forecasted DDD/1000p trends of each antibiotic to 2025 with 95% confidence intervals (CIs).An interrupted time-series on 1 January 2014 was also performed and compared to the non-interrupted time-series to evaluate the impact of AURA.Error was assessed using cross validation with an initial training period from 2000-2004, a period of 180 days, and a horizon of 360 days.This calculated the mean absolute percentage error (MAPE), which was used to assess each model's ability to model AU.

RESULTS
Antipseudomonal prescriptions had an observed increase from 2000-2012, displayed high volatility between 2013 and late 2015, and have been on a decreasing trend since 2016 (Figure 1).The mean DDD/1000p and MAPE was 5.3172 DDD/1000p and ~8.497% respectively.The highest discrepancy between observed and predicted data was on 1 April 2014, where observed levels were 175% higher than predicted (95% CI 152.16% to 207.19%).Before 2012, there were only four irregularities outside of the CI, the largest of which occurred on 1 December 2011, where the observed DDD/1000p was 23.3% lower than predicted (95% CI À8.05% to À52.38%).The increasing trend up to 2011 is thus quite indicative of the actual prescription data, while in 2012-2015 the observed prescriptions were significantly higher than predicted, with 14 values outside the 95% CI.From 2013 onwards, there was only one irregularity on 1 June 2020, in which the observed DDD/1000p was 26.73% lower than the predicted value of 3.894 DDD/1000p (95% CI À12.98% to 59.28%).From 1 January 2020, prescriptions are forecasted to decrease by 11.7% over the next 5 years to 1 January 2025 (95% CI À30.6% to 7.3%), from 5.069 to 4.478 DDD/1000p (95% CI 3.517 to 5.441).
The total DDD/1000p model was identified through combining the usage of all four antibiotics.The three listed time brackets of 2000-2011, 2012-2015, and 2016-2020 were used to analyse the trends observed.The mean DDD/1000p and MAPE values for gentamicin, tobramycin, cefepime, and ciprofloxacin were: 0.376 DDD/1000p and 10.55%; 0.8869 DDD/1000p and 242.89%; 0.196 DDD/1000p and 38.83%; and 3.856 DDD/1000p and 8.61% (Table 1) respectively.Analysis of tobramycin has shown a non-linear increasing trend (Figure 2).From 2000 to 2020, prescriptions in DDD/1000p increased from 0.02084 DDD/1000p in 2000 to 1.8605 DDD/1000p, resulting in an increase of 1.840 DDD/1000p and, according to the trend, are forecasted to increase by 0.006% (95% CI À14.7% to 12.2%).This trend is influenced by multiple irregularities above the 95% CI between 2012-2016.The most irregular data point occurred on 1 May 2015, in which the observed value was 63.8% larger than predicted (95% CI 36.4% to 107.5%).From 2016 to 2020, most observed values were also below the predicted trendline.Tobramycin accounts for 16.7% of the total mean DDD/1000p.
The time series of cefepime indicates a decreasing trend in prescriptions and can be seen as cyclical; between 2000 and 2006, observed values were higher than the trend and between 2006 and 2012, observed  prescriptions were lower than predicted (Figure 2).This cyclical nature increases the MAPE.From 2000 to 2020 prescriptions decreased by 42.2% and are predicted to further decrease 7.2% by 2025 (95% CI À69.3% to 52.6%).The dates 1 May 2006 and 1 June 2015 represent the two peaks of the evident cycle in cefepime prescriptions, and the observed DDD/1000p levels were 114.1% and 120.8% higher than the predicted values, (95% CI 41.4% to 423.5%) and (95% CI 50.4% to 265.6%) respectively.Cefepime accounts for 3.7% of total prescriptions.The time series for ciprofloxacin shows a slight increase in prescriptions from 2000 to 2010, yet an overall decrease from 2000 to 2020 (Figure 2).From 2000 to 2011, prescriptions increased by 21.4%, and from 2011 to 2020, prescriptions decreased by 39.0%.According to the trend, ciprofloxacin prescriptions are forecast to further decrease 35.98% by 2025 (95% CI À32.15% to À39.88%).Out of the 25 irregularities, 19 occurred between 2012 and 2016, mirroring the volatility seen in the total DDD/1000p time series.Ciprofloxacin accounts for 72.5% of total prescriptions.
An interrupted time-series analysis was also performed on the total prescription data to assess the impact of AURA on prescriptions of the listed antipseudomonals (Figure 3).
The mean prescriptions between 1 January 2000 and 1 December 2013 was 4.94 DDD/1000p and exhibited the lowest MAPE recorded of 3.34%.There were still several irregularities dispersed relatively evenly across the dataset, yet with comparatively low error.The highest disparity occurred on 1 October 2008, where the observed value of 6.23 DDD/1000p was 17.6% larger than the predicted prescription output (95% CI, 8.9% to 27.7%).The interrupted time-series trend forecasted an increase of 5.1% in DDD/1000p between 2014 and 2020 (95%CI, 3.0% to 13.2%).This is significantly higher than the observed decrease of 14.3% in the total data over that period (95% CI, À30.4% to 2.3%) (Figure 3).

DISCUSSION
High resistance in P. aeruginosa remains an issue of public health concern globally, with those infected associated with increased morbidity. 5s multidrug-resistant P. aeruginosa is of increasing difficulty to treat due to low antibiotic efficacy, surveillance of AMR and AU trends is essential in the selection of appropriate antipseudomonal agents for empirical therapy.Of the antibiotics studied, ciprofloxacin is of high concern considering its broad use clinically and its ability to cause outbreaks of resistant pseudomonal infections. 7Ciprofloxacin, along with gentamicin, are not only the most frequently used fluoroquinolone and aminoglycoside antibacterial agents respectively but also displayed some of the highest resistance rates in 2019 of 6.6% and 4.2% respectively 18 in Australia.The limitation in these data, that they are collected via a voluntary system, should be acknowledged.However, despite the voluntary nature of the system, participation is ever increasing. 18ntibiotic use of gentamicin, tobramycin, cefepime, and ciprofloxacin has been on the decline since 2014, after displaying an increasing trend from 2000, in part due to the introduction of AURA.This initial increase in AU is supported in past studies that surveyed a 35% global increase in antibiotic consumption between 2000 and 2010. 19The 2019 AURA report provided additional support that AU was decreasing in 2016-2017 after a steady increase in 2013. 9Little to no seasonality difference was observed in AU, despite studies suggesting AU increases in the winter months. 20However, these studies indicate seasonality was specifically driven by increased penicillin and macrolide usage, which was beyond the scope of this investigation. 20The overall decline in AU is indicative of an increase in the judicious use of antipseudomonal antibiotics.However, further studies investigating other antipseudomonal antibiotics not included in the current study are recommended to fully elucidate the current AU for pseudomonal infections in Australia.Resistance in P. aeruginosa is known to have a significant correlation with the consumption of all antibiotics, and excessive AU accelerates the development of multi-drug-resistant strains. 3Hence, the apparent decreasing trend in AU is promising in the context of P. aeruginosa.
The time series analysis of the use of gentamicin, an aminoglycoside antibiotic, indicated a linear decline from 2000 onwards.Compared to the other antibiotics, gentamicin was unaffected by AURA's introduction in 2014 and maintained a stable decline over the last 22 years.This corroborates studies suggesting aminoglycosides may be the only antibiotics showing decreasing rates of both usage and resistance. 21Studies also show decreasing usage of gentamicin has strong correlations with decreasing gentamicin resistance specifically in P. aeruginosa, 21 and these studies are consistent with AURA reports. 7he time series analysis of the use of tobramycin, another commonly used aminoglycoside antibiotic, has shown a significant increase in community AU from 2010 to 2015.Due to the high irregularity in these years, the trend in tobramycin AU from 2014 onwards cannot be deduced from the 95% CI.This sharp rise can be explained through the introduction of a dry powder solution of tobramycin by the Pharmaceutical Benefits Advisory Committee (PBAC) on March 2011. 22his administration method had a lower treatment burden with faster and more efficient pulmonary delivery than predecessors. 23This trend is somewhat consistent with the increase in tobramycin AU in the 2016 AURA review; however, the analysis of all dosage forms of tobramycin may have indicated a more accurate trend. 7he time series analysis of the use of ciprofloxacin, an oral fluoroquinolone, indicates a gradual rise in community AU from 2000 to 2010 and an overall decline from 2010 onwards.In 2013 and 2014, high MAPE was recorded due to volatile AU rates, although from 2014, AURA contributed to a stable decline in ciprofloxacin usage.The 2021 AURA review confirms this trend in AU, with upward trending resistance between 2003 and 2008, followed by a gradual decline in AMR. 10 The time-series analysis of the use of cefepime, a fourth-generation cephalosporin, indicates a gradual decline in community AU from 2000 onwards.This decline had an observed cyclical feature with a period of 6 years; however, from 2014 onwards, community AU has trended downwards, as opposed to an AU increase in 2020 that its cyclical nature suggested.Studies suggest cefepime has a low propensity towards the development of resistance, which may suggest a proportionally lower decrease in AMR compared to the other studied antibiotics. 24In countries comparable to Australia, however, cefepime only accounted for 0.01% of total cephalosporin usage. 25Hence its trends cannot be extrapolated to total cephalosporin use in pseudomonal infections.Although cefepime is listed on the PBS for febrile neutropeniaa condition that is associated with a range of infectious organismsrecent studies have shown that P. aeruginosa is a common causative organism in febrile neutropenia, causing approximately 30% of cases, 26 and is still an antibiotic of choice in the treatment of P. aeruginosa infections. 27This provides a general evaluation of the trend of AU of cefepime.However, further studies expanding the dataset to include non-PBS data are needed to fully evaluate the AU of cefepime in pseudomonal infections.
Antibiotic usage was measured through available PBS prescription data, which present limitations in interpreting the data.These data do not fully represent total community and hospital use due to the supply of non-PBS-funded medicines.However, the source of data is consistent over time and can provide data to evaluate trends in antibiotic use.Future studies investigating the use of antipseudomonal antibiotics in hospitals should be undertaken to compare trends.Additionally, future studies investigating the impacts of the 2019 novel coronavirus (COVID-19) pandemic and associated travel restrictions on AU for pseudomonal infections are recommended.9][30] In Australia specifically, community prescriptions for respiratory tract infections had significantly decreased, while those for non-respiratory infections remained unchanged early in the pandemic, 29 highlighting the need for further investigations.The data from this preliminary study suggest that, based on the four indicator antibiotics, AU in P. aeruginosa infections is declining and is forecasted to continue decreasing to 2025, potentially due to the introduction of AURA.This trend should be investigated further in future studies that include AU from a wider range of antipseudomonal antibiotics and settings, including both community and hospital usage.This could allow for a more efficient and cost-effective approach to AU and AMR surveillance in the future.Additionally, the accuracy of the forecasted trends compared to observed usage in this study suggests there is further potential application of time series analysis measuring AU.

Figure 3
Figure 3 Interrupted time series of (A) actual and (B) predicted total DDD/1000p for antipseudomonal antibiotics from 1 January 2000 to 1 January 2014 accounting for the intervention of AURA and (C) DDD/1000p forecast to 2020, including trendline, seasonality, and (D and E) mean absolute percentage error (MAPE).

Table 1
Summary of antipseudomonal antibiotic drug use as predicted using time series analysis to 2025, represented as the DDD/1000p and MAPE, for gentamicin, tobramycin, cefepime, ciprofloxacin, and total use