Development and Prospective Validation of a Machine Learning Tool for Predicting Multidrug-resistant Organism Infections in an Indian Critical Care Unit

Abstract

Background and aims

Multidrug-resistant organisms (MDROs) are a growing threat in critical care settings, causing prolonged hospitalization, increased costs, and mortality. Irrational antibiotic use, invasive procedures, and limited diagnostic times contribute to the prevalence of MDROs. Machine learning (ML) offers a promising approach for early risk identification. This study aimed to develop and pilot validate an ML-based tool to predict MDRO infection risk in Critical Care Unit (CCU) patients using baseline clinical data.

Patients and methods

Retrospective data from 323 patients admitted to the CCU of a tertiary care hospital in Kerala, India, were used to develop and evaluate five ML models using Python (version 3.11). Predictor variables included demographics, comorbidities, and initial medical device usage. A pilot prospective validation was subsequently conducted on 49 new CCU admissions to assess real-world predictive performance within 48 hours of admission.

Results

Among 323 patients, 76 developed MDRO infections (23.5%). Major risk factors identified included prolonged hospitalization and invasive device use. The random forest model demonstrated a training accuracy of 96.0% and a test-set accuracy of 75.3%. In the pilot prospective validation, the tool yielded a sensitivity of 81.8%, a specificity of 100%, and an overall accuracy of 95.9%.

Conclusions

This study demonstrates that the ML-based tool can effectively stratify MDRO risk in a CCU setting. While the performance discrepancy between training and testing indicates overfitting, the high specificity and sensitivity observed in the prospective pilot phase support its potential as a bedside aid for early antimicrobial stewardship (AMS).

Clinical significance

The ML tool enables early risk stratification, allowing for timely infection control and efficient resource utilization in constrained clinical settings.

Highlights

A machine learning (ML) tool for early multidrug-resistant organisms (MDROs) risk stratification was developed using retrospective clinical data and pilot-validated in an Indian Critical Care Unit (CCU), demonstrating significant potential for antimicrobial stewardship (AMS) within 24–48 hours of admission.

Introduction

Infectious diseases impose a substantial global health burden, presenting complex epidemiological and socioeconomic challenges that strain healthcare systems worldwide.¹ Antimicrobial resistance (AMR) is a critical component of this threat, currently associated with approximately 1.1 million deaths annually—a figure projected to reach 1.9 million by 2050.² India serves as a global epicenter for emerging infections, contributing nearly 33.9% to the worldwide prevalence due to resource constraints and socioeconomic disparities.³ Multidrug-resistant organisms (MDROs), encompassing various gram-negative bacilli and gram-positive cocci, are defined as organisms demonstrating non-susceptibility to at least one agent in three or more antimicrobial categories.⁴ These pathogens are driven by risk factors such as advanced age, prior colonization, and existing comorbidities.⁵ Critical Care Units (CCUs) are primary hubs for MDRO emergence, driven by high patient density, irrational antibiotic use, and the frequent requirement for invasive medical devices.⁶ Consequently, accurate diagnosis is essential for judicious antimicrobial selection, which significantly reduces the morbidity and mortality rates associated with MDRO infections.⁷

Machine learning (ML) offers a transformative shift in clinical diagnostics through early disease identification and real-time pattern recognition.^8,9 This technology overcomes the limitations of outdated, one-size-fits-all treatment methods by enabling tailor-made interventions optimized for individual patient profiles. Machine learning algorithms address the 48- to 72-hour diagnostic lag of traditional culture methods by analyzing vast datasets, including vitals, medical history, and laboratory results, to provide real-time risk stratification.¹⁰ Despite this potential, there is a notable lack of validated ML tools specifically developed for the unique demographic and clinical variables of Indian critical care settings. Therefore, this study aimed to develop and prospectively pilot validate an ML-based tool to predict MDRO risk specifically within the first 24–48 hours of CCU admission in a tertiary care setting in Kerala, India. The objective was to provide early risk stratification based on baseline clinical data available at the point of entry, addressing regional epidemiological variations.

Patients and Methods

This study was a two-phase cohort analysis conducted in the CCU of M.O.S.C Medical College Hospital, Kolenchery, India. The research was structured according to the Population, Intervention, Comparison, and Outcome (PICO) framework: The Population consisted of adult CCU patients (age >18 years); the Intervention was the application of an ML-based prediction tool; the Comparison was against microbiological cultures; and the Outcome was the occurrence of laboratory-confirmed MDRO infection. Reporting of this study adheres to the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD; Type 3: Development and Transformation) statement and Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines to ensure the transparent reporting of predictive model development and the systematic presentation of observational cohort data.¹¹ The Institutional Ethics Committee (IEC) of the hospital reviewed and authorized this study (Reference Number: MOSC/IEC/178/2024). All procedures were carried out following the ethical tenets outlined in the Declaration of Helsinki. For the retrospective phase (n = 323), a waiver of informed consent was granted by the IEC, as the research involved the analysis of de-identified, routinely collected clinical data without direct patient contact. Throughout the study, patient confidentiality was rigorously protected; all clinical data were anonymized at the point of collection and stored within encrypted, password-protected electronic files accessible to the primary investigators only.

Following ethical approval, a consecutive sampling method was employed, involving a total census of 323 patients admitted to the CCU over 6 months (August 2024–January 2025) who met the inclusion criteria of a CCU stay exceeding 48 hours and demonstrated bacterial growth. We excluded pregnant women, patients with incomplete medical records, and those with suspected infections who did not demonstrate bacterial growth in microbiological cultures. To ensure the model's real-time clinical applicability, predictor variables were strictly restricted to baseline clinical data available within the first 24 hours of admission. These encompassed sociodemographic details, history of hospitalizations, history of antibiotic exposure before admission, comorbidities, history of chronic kidney disease (CKD) or hemodialysis, and the presence of indwelling devices. In the absence of formal scoring systems [e.g., Acute Physiology and Chronic Health Evaluation II—(APACHE II)], age and the presence of chronic comorbidities were utilized as proxies for the severity of illness. These variables were used in accordance with those used in prior studies.^5,8,12–15

The primary outcome was the development of a laboratory-confirmed MDRO infection, defined as non-susceptibility to at least one agent in three or more antimicrobial categories.⁴ Antimicrobial susceptibility testing (AST) was performed in accordance with Clinical and Laboratory Standards Institute (CLSI) guidelines. To ensure the integrity of the assessment and minimize bias, laboratory personnel conducting the microbiological evaluations were blinded to the ML tool's risk predictions during the prospective phase. Patient characteristics were summarized using frequencies and percentages for categorical data, while continuous variables were expressed as medians with associated ranges.

Data preparation and model training were carried out using the Scikit-learn package in a Python 3.11 environment. To prepare the raw clinical data for analysis, we standardized label categories, removed duplicate entries, and conducted outlier detection as part of a multi-stage workflow. Missing data (representing <5% of the total dataset) were addressed via mean imputation to maintain the integrity of the cohort. To facilitate model training and internal validation, the retrospective dataset (n = 323) was partitioned into a stratified 70:30 train-test split to maintain the 23.5% (76/323) MDRO prevalence across subsets. Five ML models—decision trees, random forests, support vector machine (SVM), logistic regression, and Extreme Gradient Boosting (XGBoost)—were iteratively evaluated using stratified five-fold cross-validation and feature importance scaling to identify the primary drivers of MDRO risk while ensuring model stability. The highest-performing model was deployed as a real-time clinical decision support tool via the Streamlit web framework, providing instantaneous risk stratifications based on user-entered patient data. Model performance was evaluated using balanced accuracy, area under the receiver operating characteristic (AUROC) curve, sensitivity, specificity, and F1-score, with all metrics reported alongside their 95% confidence intervals (CIs).

To evaluate real-world efficacy, an initial prospective pilot validation was conducted on 49 new CCU admissions during February 2025. Risk predictions were generated for each new admission using at-admission baseline data only and cross-referenced with subsequent laboratory-confirmed microbiological culture reports. This phase served to verify the tool's predictive accuracy and potential clinical impact in a live CCU environment, providing a foundation for measuring the success of future antimicrobial stewardship (AMS) initiatives supported by the tool.

Results

To address the clinical challenge of early MDRO detection, the primary objective of this study was to develop and validate an ML model to determine whether baseline clinical parameters available at admission can effectively predict the subsequent occurrence of these infections in the CCU. We analyzed a retrospective cohort of 323 patients admitted to the unit. The cohort consisted of 216 males (66.9%) and 107 females (33.1%), with an overall mean age of 59.54 ± 16.5 years. Laboratory evaluation confirmed positive cases in 76 patients, establishing a prevalence of 23.5%. Demographic stratification (Table 1) revealed that infection rates were notably higher in female patients (29.9%) compared with males (20.4%). Furthermore, antimicrobial susceptibility showed a significant correlation with advancing age, with the highest prevalence observed in patients over 60 years (28.4%). All 323 patients were systematically accounted for across demographic and admission categories to provide a comprehensive baseline analysis for the model development cohort.

Table 1.

Clinical and demographic risk factor distribution (N = 323)

Variable	Total (N)	MDRO-positive (n = 76)	Prevalence (%)
Gender
Female	107	32	29.9
Male	216	44	20.3
Age
>60 years	169	48	28.4
<60 years	154	28	18.1
Length of stay
>15 days	60	28	46.6
10–15 days	56	13	23.2
5–10 days	146	30	20.5
≤5 days	61	5	8.2
History (past 3 months)
Antibiotic exposure
Prior antibiotic exposure	56	25	44.6
No prior antibiotic exposure	267	51	19.1
Hospitalization
Prior hospitalization	75	27	36.00
No prior hospitalization	248	49	19.7
Type of admission
Direct admission	232	57	24.5
Transfer from another hospital	91	19	20.8

%, Percentage; MDRO, multidrug-resistant organism; n, number of patients in a subgroup; N, total number of patients

Microbiological evaluation confirmed that Escherichia coli, Klebsiella pneumoniae, Acinetobacter baumannii, and Pseudomonas aeruginosa were the predominant isolates, typically sourced from urine, sputum, blood, and endotracheal secretions. The distribution of resistance across antibiotic categories revealed a significant clinical burden, with the highest frequency of resistance observed in the Penicillin class (n = 148), followed by cephalosporins (n = 97), fluoroquinolones (n = 69), and aminoglycosides (n = 48), as shown in Supplementary Table 1. Lower resistance frequencies were noted for sulfonamides (n = 30), carbapenems (n = 27), macrolides (n = 11), and nitrofurans (n = 10), reflecting the selective pressure exerted by broad-spectrum agent usage within the unit.

Analysis of clinical characteristics within the retrospective cohort identified a notable association between the total duration of hospitalization and infection prevalence. Patients with a stay exceeding 15 days exhibited an infection rate of 46.6%, compared to 20.6% in the 5- to 10-day group (Table 1). Similarly, clinical history served as a descriptive marker for risk, with higher frequencies observed in subjects with prior admissions (36.0%) or antibiotic exposure (44.6%) within the preceding 3 months. Regarding admission types, medical emergencies (24.0%) and direct admissions (25.6%) showed slightly higher infection rates compared with surgical emergencies (21.7%) or transfers (20.9%). All categories were calculated based on the total census of 323 patients to ensure consistency across the reported data. To ensure methodological alignment between both study phases, these retrospective outcomes were utilized solely for cohort characterization, while the features selected for model development were strictly restricted to baseline parameters available at the time of admission, mirroring the data entry protocol utilized in the subsequent prospective validation.

A comprehensive analysis of comorbidities and medical devices (Table 2) identified several high-prevalence subgroups. While certain rare conditions showed high infection frequencies—such as seizures (100%) and rheumatoid arthritis (50%)—these findings were notably limited by small sample sizes within those specific subsets. More statistically representative risk factors included lower respiratory tract infections (66.7%) and CKD requiring hemodialysis (62.5%). The presence of invasive medical devices was consistently associated with increased risk, including non-invasive ventilation (27.1%), Ryle's tube (25.7%), Foley's catheterization (23.7%), and central venous lines (23.6%). Additionally, postoperative status (35.7%) and a general history of CKD or dialysis (32.6%) were identified as prominent clinical indicators within the cohort.

Table 2.

MDRO infection prevalence by comorbidity and medical device (N = 323)

Variable category	MDRO-positive (n)	Prevalence (%)
Comorbidities
Seizure	1	100.0
LRTI	4	66.6
CKD with hemodialysis	5	62.5
Rheumatoid arthritis	1	50.0
CVA	1	50.0
BPH	2	40.0
T2DM	48	32.2
Hypothyroidism	3	27.2
Hypertension	37	25.1
Asthma	1	25.0
Medical devices
Non-invasive ventilation	29	27.1
Ryle's tube	28	25.6
Foley's catheterization	66	23.7
Central line	17	23.6

%, Percentage; BPH, benign prostatic hyperplasia; CKD, chronic kidney disease; CVA, cerebrovascular accident; LRTI, lower respiratory tract infection; MDRO, multidrug-resistant organism; n, number of patients in a subgroup; N, total number of patients; T2DM, type 2 diabetes mellitus

To develop the predictive tool, five ML algorithms—including decision trees, random forests, SVM, XGBoost, and logistic regression—were evaluated using stratified five-fold cross-validation. Preliminary analysis revealed that decision trees yielded suboptimal predictive performance and were consequently excluded from final model optimization. Performance analysis indicated that SVM and logistic regression achieved the lowest test accuracies of 67% and 68%, respectively. While training accuracies reached 96% for the random forest model, performance on the unseen test set (n = 97) demonstrated a more generalized accuracy of 75.3% (95% CI: 65.4–83.5; Fig. 1), outperforming the XGBoost model. Feature importance scaling (Fig. 2) identified the history of antibiotic exposure before admission as the most influential baseline predictor, followed by comorbidity profiles and the initial requirement for indwelling devices. The model's overall performance was quantified by a binary area under the curve (AUC) of 0.96 (Fig. 3); this metric specifically reflects the tool's capacity to discriminate between the two primary study outcomes, MDRO-positive and MDRO-negative, based on the aggregated baseline feature set.

Fig. 1.

Comparison of ML models’ training and testing accuracies

Fig. 2.

Feature importance plot for the random forest model

Fig. 3.

ROC curve of the model

To evaluate the real-world efficacy of the model, a prospective observational pilot study was conducted on a new cohort of 49 CCU admissions. Upon admission, patient variables were input into a custom Streamlit-based web application (Supplementary Fig. 1) to classify individuals as “high risk” or “low risk.” The tool identified nine patients as “high risk” upon admission, while subsequent clinical and microbiological testing confirmed 11 total cases of MDRO development. Cross-referencing revealed that all nine patients flagged as “high risk” by the tool were among the 11 confirmed cases, representing a sensitivity of 81.8% (95% CI: 52.3–94.9). Notably, the model generated zero false positives in this pilot phase, resulting in a precision (positive predictive value) and specificity of 100% (95% CI: 90.8–100.0). Overall, the tool correctly classified 47 out of 49 patients, yielding a prospective accuracy of 95.9% (95% CI: 86.3–98.9) and an F1-score of 0.90. While these high-performance metrics demonstrate strong clinical potential, the reported specificity and 95% CIs should be interpreted within the context of the small pilot sample size (N = 49), which reflects the preliminary nature of this real-time validation (Table 3).

Table 3.

Confusion matrix for prospective validation (n = 49)

Actual status	Predicted MDRO (+)	Predicted MDRO (−)
Actual MDRO (+)	9 (true positive)	2 (false negative)
Actual MDRO (−)	0 (false positive)	38 (true negative)

Performance Metrics: Accuracy: 95.9% (95% CI: 86.3–98.9); sensitivity: 81.8% (95% CI: 52.3–94.9); specificity: 100.0% (95% CI: 90.8–100.0). (−), negative; (+), positive; MDRO, Multidrug-resistant organism; n, number of patients

Discussion

This study was conducted in the CCU of a tertiary care hospital with the primary aim of developing and pilot-validating an ML-based prediction tool for the early identification of patients at high risk for MDRO infections. The results revealed a substantial burden of infections among critically ill patients, with a prevalence rate of 23.5% in the development cohort of 323 patients. This finding is consistent with existing literature that emphasizes the high incidence of MDRO infections in critical care settings due to the frequent use of invasive devices, prolonged hospital stays, and intensive antibiotic exposure. Microbiological analysis identified E. coli, K. pneumoniae, A. baumannii, and P. aeruginosa as the predominant isolates. Escherichia coli and K. pneumoniae were the most frequent isolates in urine and sputum samples, aligning with previous studies that identify these bacteria as primary culprits in catheter-associated urinary tract infection (CAUTI) and ventilator-associated pneumonia (VAP). Karukappadath et al. provide a comprehensive study regarding drug-resistant bacteria in the critically ill and highlight the “ESKAPE” pathogens—Enterococcus faeciu m, Staphylococcus aureus, K. pneumoniae, A. baumannii, P. aeruginosa, and Enterobacter species—as particularly problematic due to their sophisticated resistance mechanisms and ability to evade standard treatments.¹⁶

The ML model incorporated a wide range of risk factors that significantly correlated with MDRO infection risk. Vasudevan et al. conducted a prospective observational study on multidrug-resistant gram-negative bacilli (MDRGNB) and identified mechanical ventilation (p = 0.01), central line application (p = 0.05), urinary catheterization (p = 0.04), antibiotic exposure within the past 6 months (p = 0.04), and end-stage renal disease with dialysis (p = 0.017) as major risk factors.⁵ Our findings mirror these results, as baseline parameters available upon admission, such as initial mechanical ventilation and indwelling device requirements, were the primary predictors in our random forest model. By focusing on these specific features, the model maintained high discriminative power despite the inherent class imbalance, where MDRO-positive cases constituted 23.5% of the cohort, ensuring that the algorithm accurately identified high-risk subjects without over-relying on the majority non-infected group.

The statistical significance of these factors is further supported by Chen et al., whose study on nosocomial infections showed that multidrug resistance and prior antibiotic exposure remained significant clinical indicators (p < 0.05).¹³ Furthermore, Wang et al. conducted a retrospective study among the critically ill and demonstrated that factors with high statistical significance (p < 0.001) included past antibiotic exposure and invasive procedures before CCU admission, and APACHE II scores. Their study also confirmed the significance of central line use (p = 0.007) and other devices like non-invasive ventilation (p = 0.013), all of which were integral features in our predictive framework.¹² While scoring systems like APACHE II measure overall disease severity, they lack pathogen-specific prediction; conversely, our model utilizes baseline clinical data for targeted AMR risk assessment. We distinguished retrospective characteristics used for cohort characterization, such as total length of stay (p < 0.01), from the prospective features utilized in the predictive tool. By restricting inputs to data identifiable within the first 24 hours of admission, the tool provides true early-warning stratification. This ensures results represent a proactive baseline assessment to support AMS, rather than a reflection of post-infection outcomes.

Antibiotic resistance profiling in our study revealed a concerning trend of resistance to widely used broad-spectrum agents, including penicillins (n = 148), cephalosporins (n = 97), and fluoroquinolones (n = 69). These patterns indicate that standard empiric therapies may be increasingly ineffective in critical care settings. In a study by Ramsamy et al., a high prevalence of resistance to these same drug classes was found among E. coli, K. pneumoniae, and gram-positive organisms such as Staphylococcus species, highlighting the urgent need for individualized, evidence-driven antimicrobial strategies.¹⁷ This necessity is further underscored by Rawson et al., who demonstrated that supervised ML can leverage routine baseline data to predict clinical infection profiles upon hospital admission.¹⁸ Our ML model addresses this clinical gap by providing a targeted risk assessment that supports AMS goals. By identifying high-risk patients at the point of admission, the tool enables a more precise transition from broad-spectrum empiric coverage to individualized therapy, potentially mitigating the further selection of resistant strains in the CCU environment.

The development of this tool serves as a transformative approach to real-time clinical decision-making. Of the models evaluated, random forest provided the highest predictive performance, echoing Fleuren et al. regarding the utility of such models for rapid stratification in high-stakes clinical conditions.¹⁹ Although training accuracy reached 96.0%, the 75.3% accuracy on the unseen test set suggests a degree of overfitting, a common challenge where models over-index on specific retrospective cohort patterns. However, the random forest architecture maintained clinical utility by aggregating predictions across multiple decision trees. The model's strong discriminative capacity was further evidenced by an AUC of 0.96 (Fig. 3), reflecting a robust ability to distinguish between MDRO-positive and MDRO-negative cases. This was confirmed in a 49-patient prospective pilot, where the tool identified 9 of 11 MDRO cases [81.8% sensitivity (95% CI: 52.3–94.9)] with 100.0% specificity (95% CI: 90.8–100.0) and 95.9% accuracy (95% CI: 86.3–98.9). While potentially reflective of this specific cohort, the high positive predictive value suggests significant potential to minimize unnecessary resource allocation and isolation protocols, provided these results are further validated in larger cohorts.

This study presents a site-specific ML tool specifically developed to address the unique epidemiological patterns and clinical characteristics of an Indian CCU. The inclusion of a pilot prospective validation phase in a live clinical environment demonstrates the tool's practical utility for real-time risk stratification beyond purely retrospective datasets. Furthermore, by strictly restricting model inputs to baseline clinical data identifiable within the first 24 hours of admission, the tool offers a proactive, early-warning assessment rather than a reflection of post-infection outcomes. Finally, the deployment of a custom Streamlit web application provides an accessible, user-friendly interface designed for real-time bedside risk stratification and prospective AMS support.

Limitations

Conversely, several limitations must be acknowledged. Primarily, data were collected from a single tertiary care hospital, which predominantly reflects local ecological trends and resistance patterns, thereby limiting the broader generalizability of the model. Additionally, the model exhibited clear evidence of overfitting, indicated by the performance gap between training (96.0%) and test-set (75.3%) accuracy, which remains a frequent methodological constraint in retrospective development cohorts. The presence of class imbalance (23.5% MDRO prevalence) in the development data also necessitates a cautious interpretation of standard accuracy metrics. Furthermore, while the retrospective training phase was substantial, the prospective validation phase was pilot scale (N = 49), which restricted the overall sample size and limited the long-term monitoring of clinical outcomes. Another constraint is that despite restricting variables to the first 24 hours of admission, potential temporal data leakage remains a concern for certain retrospective predictor variables. Moreover, while this preliminary evaluation focused strictly on model discrimination (AUROC, sensitivity, and specificity), formal reporting of calibration metrics—such as the Brier score and calibration plots—was restricted by the pilot scale of the prospective cohort and remains an essential requirement for future multi-center external validation studies. Incomplete documentation in medical records also necessitated the use of proxy markers, such as age and chronic comorbidities, for illness severity instead of formal scoring systems like APACHE II. Finally, direct clinical impacts, such as changes in mortality rates or resource-allocation efficiency, were not formally measured in this preliminary pilot phase.

Conclusion

This study successfully developed and pilot-validated an ML framework for identifying MDRO risk at the point of CCU admission. By leveraging a core set of baseline risk factors, including prior antibiotic exposure, chronic comorbidities, and initial indwelling device requirements, the random forest model effectively stratified patients into risk categories within the first 24 hours of admission. In the prospective pilot phase, the tool yielded an accuracy of 95.9% and a sensitivity of 81.8% without generating false positives in this specific cohort.

Although the 75.3% test accuracy indicates a degree of model overfitting due to the preliminary nature of the pilot sample size, the framework demonstrates the clinical viability of baseline clinical data for targeted risk assessment before traditional microbiological data is available. However, before this model can be integrated into standard clinical workflows, external validation in larger, multi-center cohorts is essential to confirm its generalizability and real-world deployment readiness.

Clinical Significance

The clinical significance of this study lies in its potential to enable targeted interventions at the point of CCU admission, effectively bridging the critical diagnostic gap before microbiological culture results are available. By providing early risk stratification, the ML-based tool allows clinicians to ensure that intensive resources, such as contact isolation protocols and potent reserve antibiotics, are allocated more efficiently based on individual patient risk profiles.

This proactive strategy directly supports AMS programs by reducing the traditional reliance on broad-spectrum empiric therapy and minimizing the unnecessary use of high-tier agents. Furthermore, the successful pilot validation in a resource-limited setting underscores the model's potential as a scalable bedside decision support tool. This framework offers a preliminary path toward improving patient care by streamlining infection control workflows and significantly reducing delays in initiating targeted therapy during the first 48 hours of hospitalization.

Acknowledgments

The authors would like to express their sincere gratitude to the Departments of Critical Care and Microbiology, along with the Institutional Ethics Committee at MOSC Medical College Hospital, Kolenchery, for their administrative support and for providing access to the clinical data essential for this study. They are equally thankful to the management and faculty of Nirmala College of Pharmacy, Muvattupuzha, for their academic guidance throughout this research, and they acknowledge Angel Sara Peter for providing technical assistance in the development of the ML algorithms and the web application.

Declarations

Ethical Approval

The Institutional Ethics Committee (reference number-MOSC/IEC/178/2024) approved the research protocol.

Consent for Publication

Not applicable. This study involved the analysis of de-identified, retrospective clinical data, and a waiver of informed consent was granted by the Institutional Ethics Committee.

Clinical Trial Registry Number

Not applicable, as this was an observational tool development study; a formal study protocol was not prospectively registered for this analysis.

Sources of Funding

None.

Conflict of Interest

The authors declare no conflicts of interest.

AI Disclosure Statement

The authors acknowledge the use of Gemini (Google) exclusively for grammar enhancement and language editing of the manuscript text. No artificial intelligence (AI)-assisted technologies were employed for data collection, software coding, or statistical analysis, and all suggested edits were critically reviewed by the authors to ensure clinical integrity.

Authors’ Contributions

Nadheem M Shajeef: Conceptualization, methodology, software, data curation, and writing—original draft. Teresa M Sobi: Conceptualization, investigation, data curation, and writing—original draft. Chithra Jayaprakash: Validation, investigation, and resources. Suja Abraham: Project administration, supervision, and writing—review and editing.

Orcid

Nadheem M Shajeef https://orcid.org/0009-0007-9149-7866

Teresa M Sobi https://orcid.org/0009-0000-0355-5149

Chithra Jayaprakash https://orcid.org/0009-0001-8309-1045

Suja Abraham https://orcid.org/0000-0003-4085-8792

Supplementary Materials

All the supplementary materials are available on its journal website www.ijccm.org.

Data Availability Declaration

The datasets and the analytical Python code generated during the current study are not publicly available due to ethical and confidentiality restrictions, but are available from the corresponding author upon reasonable request.