Ranked Placement of Phage Predation as a Determinant of Dehydration Severity Among Cholera Patients in Bangladesh

Abstract

Virulent bacteriophages (phages) can kill bacterial prey, potentially reducing burden of infection. In cholera, a high phage to Vibrio cholerae ratio is associated with mild dehydration, yet the relative importance of this ratio in disease severity remains unclear. We used machine learning to rank select host, microbial, and environmental factors as determinants of dehydration severity in over 600 cholera patients from across Bangladesh. We found the phage:pathogen ratio ranked among the top classifiers for mild dehydration, behind age and location. We advocate that phage predation be included as a key factor in cholera characterization for scientific, clinical, and epidemiological applications.

In cholera, Vibrio cholerae is targeted by virulent bacteriophages. However, the relative importance of phage predation on disease severity is unclear. Machine learning showed that phage predation ranked within top features associated with mild dehydration.

Cholera is an acute diarrheal disease caused by the Gram-negative bacterium Vibrio cholerae (Vc). Determinants of severe dehydration include diverse host, microbial, and environmental factors [1–4]. We previously found that higher ratios of a Vc-specific phage (ICP1) to its prey (Vc) were associated with mild dehydration [1]. This suggests that ICP1 naturally suppresses Vc within infected patients, resulting in less severe symptoms. The ICP1:Vc ratio thus may have potential translational applications, for example in improved diagnostics of both infection and disease severity. Using a simple threshold of 1 ICP1 to 5 Vc, we differentiated patients with and without dehydration (85% sensitivity; 53% specificity) [1]. While intriguing, the relative importance of phage predation as a determinant of dehydration compared with known factors, such as patient age, is unknown. Addressing this knowledge gap is key to better characterize cholera pathophysiology and inform clinical, diagnostic, and public health strategies.

To address this gap, we applied multiple machine learning (ML) models to a previously collected cohort of over 600 cholera patients enrolled at hospital admission at 7 hospitals across Bangladesh. Our objective was to identify and rank phage as a potential determinant against known and unknown determinants of dehydration. We selected a short list of host demographic (age, gender, and hospital location), microbial (Vc and phage ICP1 absolute abundances, and the ICP1:Vc ratio), and environmental (antibiotic exposure determined by mass spectrometry) factors as potential predictors. Our guiding hypothesis was that phage predation would rank high in importance against known predictors in discriminating patients with and without dehydration.

METHODS

Ethics Statement

The parent study was led by A. I. K. and F. Q. at the icddr,b in Bangladesh and E. J. N. who led the laboratory research at the University of Florida. B. J. S. conducted analyses on a deidentified dataset. Associated protocol numbers/registrations were IEDCR IRB/2017/10; icddr,b ERC/RRC PR-17036; University of Florida IRB 201601762; clinicaltrials.gov NCT03154229.

Study Design

This study represents a secondary analysis of a longitudinal prospective study of patients with diarrheal disease who were admitted to 1 of 7 hospitals across Bangladesh in a 1-year period [5]. We analyzed records of the 623 cholera patients identified out of 2462 total patients enrolled.

Dataset Description

A cholera patient was defined microbiologically by culture positivity for Vc, molecular detection of Vc, or the Vc-specific bacteriophage (ICP1) [1, 5]. Among the 623 patients, 82 (13.2%) had “Mild,” 254 (40.8%) had “Moderate,” and 287 (46.1%) had “Severe” dehydration by WHO measures. Mass spectrometry (LC-MS/MS) was performed to quantify antibiotics in stool samples from a subset of cholera patients (n = 266) described previously [1]; this smaller dataset is referred to as the “subset,” in contrast to the “full” dataset of 623 patients. Of those samples in the subset dataset, 16 (6%) had “Mild,” 101 (38%) had “Moderate,” and 149 (56%) had “Severe” dehydration. Quantitative polymerase chain reaction was performed on all the samples to determine absolute abundances of Vc and phages (ICP1/2/3) [1].

Machine Learning Analyses

We considered demographic (age, gender, and hospital location), microbial (Vc and phage ICP1 absolute abundances, and the ICP1:Vc ratio), and concentrations of 3 antibiotics quantified by mass spectrometry (for the subset in which they were measured) as predictors of dehydration severity (‘mild,” “moderate,” or “severe”). Using the WHO conventions which have variable accuracy, patients with dehydration (“nonmild”) have at least 3%–5% body weight loss [6]. Analyses were conducted using 7 machine learning (ML) algorithms from the scikit-learn Python library (version 1.4.1), Python 3.9.18. These included 3 implementations of gradient boosting decision tree models: Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Categorical Boosting (Catboost) [7]. The other algorithms included the ensemble decision tree method, Random Forests (RF), the distance-based K-nearest neighbors, and 2 linear models: linear support vector machine and logistic regression. Models were trained on 80% of the data and tested on the remaining 20% holdout set, using the Python package scikit-learn 1.4.1, in Python 3.9.18. To assess performances, the classifiers were trained on each training dataset produced by stratified 100-fold cross-validation (CV) of the input dataset. See Supplementary materials for additional details on the ML approach, including hyperparameter tuning, CV, and performance metrics. In the final analysis, we identified and ranked features contributing most to the classification, we calculated their importance using the Shapley Additive Explanation (SHAP) method [8]. Shapley Additive Explanation (Shap 0.44.0 package in Python) is an approach that explores output from ML models using game theory. It provides a method to evaluate the contribution of each feature to the classifier's prediction.

RESULTS

We trained ML classifiers on input data (eg, patient characteristics, antibiotic exposures, phage:pathogen ratios) to identify and rank features predictive of the clinical outcome of disease severity (‘mild,” “moderate,” and “severe” dehydration). We assessed the ability of 7 different ML algorithms to correctly classify outcomes when faced with an unseen fraction of the data held back from the training set. Catboost and RF performed best in all performance metrics (Supplementary Table 1). Both RF and Catboost classifiers showed similar performance and achieved similarly high area under the curve (AUC) under the receiver operating characteristic curves, a metric known to be invariant to class imbalance [9, 10]. The RF model achieved a comparable performance with the Catboost model (mean AUC of 0.821; 95% CI: 0.816–0.826 for RF, which overlapped with the mean AUC of 0.825; 95% CI: 0.82–0.83 for Catboost) when trained on the full dataset (n = 623; Supplementary Table 1). These 2 models generally agreed in their ranking of feature importance. For simplicity and interpretability, we reported only SHAP feature importance results from the RF model. Unlike Catboost, the RF-based SHAP values represent each level of categorical features as separate features, showing their individual importance in the prediction.

Across all models, sampling location and age were identified as the top 2 most important predictors of dehydration severity (Figure 1). Among microbial features, the models revealed 3 key findings: (i) The ratio of phage (ICP1) to Vc was the strongest positive predictor for mild dehydration; (ii) absolute ICP1 abundance was the strongest negative predictor for moderate dehydration; (iii) absolute Vc abundance was the strongest positive predictor for severe dehydration (Figure 1).

Figure 1.

Feature importance from the RF model trained on the full dataset. Directional relationships between each patient feature and (A) mild, (B) moderate and (C) severe dehydration. D, The importance ranking of features according to the average absolute impact (absolute mean SHAP values) over all samples. Positive SHAP values indicate features that increase the prediction for nonmild status, and negative SHAP values indicate features that decrease the prediction of nonmild. Larger SHAP values indicate a greater influence on the model's prediction. Each point represents a single patient. Y-axis: features ordered by their importance in the prediction. X-axis: SHAP values indicating magnitude and direction of influence. The model was trained on 623 patients (82 mild, 254 moderate, and 287 Severe) and 6 features (Vc, ICP1, ICP1:Vc, patient's age and gender, and hospital ID). Microbial features (Vc, ICP1, and their ratio) are shown in bold.

To assess the importance of antibiotic exposures in classifying disease severity, we repeated the analyses on a subset of 266 samples with antibiotics quantified from stool. We note that this subset is enriched in culture-positive cholera cases (211/266) relative to the overall dataset (219/623). Random Forest and Catboost classifiers of dehydration severity performed similarly to the full dataset, with only slightly lower mean AUC values (Supplementary Table 1). The RF classifier achieved a slightly higher AUC (mean of 0.801), but with a 95% CI (.793–.809) that overlapped with the Catboost classifier (mean AUC of 0.795). In the RF classifier applied to the subset of samples with antibiotic exposure data, the ranking of the top predictive features was similar to the full dataset (Figure 1). Antibiotics (ciprofloxacin, doxycycline, and azithromycin) ranked relatively low in importance (Supplementary Figure 1). To confirm the relatively low importance ranking of antibiotics in classifying dehydration severity, we reran the RF classifier on the same subset (n = 266 patients) after withholding the antibiotic exposure data from the model. The classifiers run on the same subset of patients were indistinguishable in their performance, yielding a mean AUC of 0.804 including the antibiotic data (95% CI: .797–.812) and 0.801 without the antibiotic data included (95% CI: .793–.809). The relative ranking of feature importance was similar in models with (Supplementary Figure 1) or without (Supplementary Figure 2) antibiotic data included, and with different rates of culture positivity. Together, phage ICP1, Vc, and their ratio were more important determinants of dehydration severity than any of the 3 antibiotic exposures assessed by mass spectrometry.

DISCUSSION

In this study, we aimed to evaluate and rank phage predation as a determinant of dehydration severity among cholera patients in Bangladesh. To address this objective, we leveraged ML-based approaches to evaluate the relative importance of phage predation against known demographic determinants [2–4, 11–13]. This objective is important because phage predation was a previously known determinant, however we lacked insight on its relative importance [1]. Taken together, we found the phage to pathogen ratio ranked among the top predictive features of dehydration, placing just behind patient age and admission location. Despite our previously observed correlations between higher antibiotic exposures and mild dehydration severity [1], we tentatively conclude that phage predation is a more important determinant of dehydration severity. Even though antibiotic concentrations were of limited predictive value in this study, they were associated with mild dehydration and certain antibiotic resistance genes in our previous analysis of the same cohort [1]. We advocate that future studies account and control for pre-enrollment antibiotic exposures as a determinant of disease severity as well as a potential confounding factor.

After hospital location and age, the models ranked phage predation (ICP1:Vc ratios) as the best predictor for mild dehydration, ICP1 absolute abundance as the best predictor for moderate dehydration and Vc absolute abundance as the best predictor for severe dehydration, regardless of whether antibiotics were included in the models. At face value, these findings are consistent with an underlying mechanism of phage ICP1 reducing the number of viable Vc cells within an infected patient, thereby reducing the severity of disease. With more detailed time-series sampling paired with sequencing both Vc and ICP1 populations, we may gain deeper mechanistic insights into how these predator-prey dynamics arise as a cause or consequence of disease progression.

Applications of these findings may impact our fundamental understanding of the pathophysiology of cholera and related diarrheal diseases, diagnostic development, and epidemiologic surveillance. While phage predation has been documented for over a century, considering phage predation in disease frameworks may help explain why seemingly identical patients have dramatically different clinical presentations and courses. Diagnostics might be improved by using phage as a proxy for pathogen detection when the phage is pathogen-specific and degrades the pathogen [14]. Lastly, improved diagnostics will in turn improve epidemiologic surveillance. Future validation, including cost analyses, is needed before implementation. For example, PCR-based detection of ICP1 could be added at relatively low cost to existing PCR-based assays for Vc. Such a dual PCR assay could potentially identify true cholera patients that would have been missed due to ICP1 pushing Vc below the limit of detection, while also yielding information about disease severity. Antigen-based rapid diagnostic tests for phage are also under development [15], which could further reduce costs while yielding rapid results.

These findings need to be considered in the context of the study limitations. First, the sample size was relatively unbalanced, skewing toward more severe dehydration. To account for this imbalance, we used cost-sensitive classifiers and stratified CV, and focused on AUC as a performance metric that is relatively insensitive to class imbalance. Nevertheless, our findings should be replicated in larger, and ideally more balanced cohorts. Second, relatively few features were included in the analysis to keep the analysis focused; these features were based on prior literature yet may have biased the analysis by unknowingly excluding potentially important features. Moreover, absolute antibiotic concentration in stool may not reflect the timing and magnitude of exposure with direct and simple kinetics. Third, the samples analyzed were collected at hospital admission alone. Given that infection, phage predation, and transmission are dynamic processes, future sampling of patients over time will be informative. We have recently collected a 24-hour time series study of cholera patients, which will help determine whether ICP1:Vc ratios at admission are predictive of future disease outcomes.

CONCLUSION

This study ranks phage predation for the first time within the fold of known factors associated with dehydration severity among cholera patients. The strengths and limitations of this study should inspire further research to test our core hypotheses for both cholera and potentially other common enteric pathogens that are likely influenced by phage predation.

Supplementary Data

Supplementary materials are available at The Journal of Infectious Diseases online (http://jid.oxfordjournals.org/). Supplementary materials consist of data provided by the author that are published to benefit the reader. The posted materials are not copyedited. The contents of all supplementary data are the sole responsibility of the authors. Questions or messages regarding errors should be addressed to the author.