Urine cultures are the primary method for urinary tract infection diagnosis. Like most culturing applications, these tests often require days to yield conclusive results, causing harmful treatment delays. Additionally, hospitals waste substantial time and resources processing negative patient samples. Predicting culture outcomes before their final result can accelerate patient care and improve diagnostic efficiency by minimizing resource allocation towards culturing negative samples. The goal of this study is to build machine learning models to predict urine culture outcomes and help optimize test order protocols. Using clinical data from the UW Medical Center, containing urine culture results, blood test results, and demographics from over 88,000 patients, I trained Random Forest models, Support Vector Machines, and XGBoost models to predict overall culture positivity and specific infection types (E. coli, Klebsiella, etc.). Predictive parameters included common clinical laboratory tests (complete blood counts, metabolic panels, etc.), as well as demographics (age, sex, and race). To evaluate the predictive performance of these features at different points in the culturing timeline, the data was divided into three subsets: (1) patients with blood work up to 30 days prior to sample collection, (2) patients with blood work up to 30 days prior to their first culture result, and (3) patients with blood work predating their final culture result. Tree-based ensemble models (RF and XGBoost) trained on the latter two subsets yielded the most promising results. The Random Forest model’s AUC (area under the curve), a value between 0 and 1 that measures a model’s ability to distinguish between two classes, was 0.87. An AUC closer to 1 indicates more accurate classifications. The results show that ML models can feasibly predict culture results to optimize operations and enable earlier treatment. Further development of this data pipeline will allow for detailed predictions of specific infection types and concurrent infections.

Critical care outcomes vary significantly across healthcare settings, and resource-limited environments often lack high-quality data-driven insights that can improve patient care. In particular, there is a strong need to understand whether insights derived from high-resource settings such as the US can be readily applied to more resource-constrained healthcare settings. To this end, here I systematically compared ICU populations between hospitals in India and the University of Washington Medical Center (UWMC) to assess differences in laboratory test dynamics and clinical outcomes (mortality, sepsis, blood transfusions, etc.). Using ICU datasets containing >100,000 patients, I applied clustering algorithms to segment patient populations based on lab test time series data. Clustering was performed across stratifications of age, sex, and admitting diagnosis to capture variations in population health. Resultant clusters demonstrated up to a threefold stratification in mortality rate (3% to 15%), with similar stratifications for other outcomes. Additionally, dimensionality reduction techniques (Uniform Manifold Approximation and Projection [UMAP], t-Distributed Stochastic Neighbor Embedding [t-SNE], etc.) were used to visualize population differences. To identify the strongest predictors of clinical outcomes, I used Cox proportional hazards model to quantify the impact of individual lab tests on patient outcomes. The analysis revealed distinct differences in lab test significance between UWMC and Indian ICU populations. Key lab tests such as Blood Urea Nitrogen and Hemoglobin were strong predictors of adverse outcomes across both cohorts, but their relative importance varied between the two settings, suggesting differences in disease progression, healthcare practices, and marker utility. These findings highlight the need for region-specific risk models and emphasize the importance of integrating time-series lab data into ICU decision-making. Understanding these differences can inform the development of more generalizable risk stratification models and improve critical care strategies in resource-limited settings, ultimately advancing global healthcare equity.

Clinical ventilation methods - where a patient’s breathing is supported by a cannula, a mask, or in severe cases, tracheal intubation – are often used to support patients in ICUs (intensive care units). Given the high risks associated with intubation, a key question is whether a patient's level of care can be de-escalated. This question is crucial in the context of newborn children (or neonates) who are especially fragile.  To address this question, we developed machine learning models to predict ventilatory support outcomes in neonates. Using a multi-center dataset of 6,538 neonates from ICUs (Intensive Care Units) across 121 Indian cities, we extracted time-based snapshots of each patient’s clinical status at 6-hour, 12-hour, and 24-hour intervals. For each snapshot, we utilized features related to vitals (heart rate (HR), respiratory rate (RR), etc.), demographics (e.g., highest HR, highest RR, etc.), and ventilation states (e.g., room air, nasal cannula, noninvasive ventilation, intubation). We trained Random Forest models, with hyper-parameters optimized via 5-fold cross-validation, using a 70-30 train-test split, to predict three outcomes: (1) mortality, (2) escalation of ventilatory support (“step-up”), and (3) successful discharge. Model performance was evaluated using the area under the receiver-operating-characteristic curve (AUROC). Predicting at 6, 12, and 24 hrs, the best models achieved out-sample AUROC values of 0.79–0.87 for mortality, 0.81–0.85 for step-up, and 0.68–0.70 for discharge. The most predictive variables for mortality, escalation, and safe discharge were the highest past FIO2, and the presence of a cannula for escalation and safe discharge, respectively. These findings suggest that machine learning can provide early warnings of clinical deterioration, reducing unnecessary intubations and improving neonatal respiratory care. By enhancing risk stratification, particularly in resource-limited settings, this approach may guide more timely and judicious use of tracheal intubation and contribute to improved neonatal outcomes.