Explainable Admission-Level Predictive Modeling for Prolonged Hospital Stay in Elderly Populations: Challenges in Low- and Middle-Income Countries

TL;DR

This study tackles predicting prolonged hospital stay (pLoS) among elderly patients in a resource-limited setting. It introduces a novel feature-selection pipeline that combines Weight of Evidence (WoE), Information Value (IV), and graph-theoretic clique reduction, followed by OptimalBinning, to produce a small, highly interpretable set of 9 variables for a logistic regression predictor. On a large LMIC hospital dataset, the model achieves an AUC-ROC of $0.82$ on the validation set with strong calibration, surpassing baseline feature-selection approaches in recall while maintaining competitive accuracy. SHAP analyses corroborate the interpretability of the WoE-based results, supporting practical deployment to aid bed management and guide future intervention studies aimed at reducing pLoS in elderly hospital populations.

Abstract

Prolonged length of stay (pLoS) is a significant factor associated with the risk of adverse in-hospital events. We develop and explain a predictive model for pLos using admission-level patient and hospital administrative data. The approach includes a feature selection method by selecting non-correlated features with the highest information value. The method uses features weights of evidence to select a representative within cliques from graph theory. The prognosis study analyzed the records from 120,354 hospital admissions at the Hospital Alma Mater de Antioquia between January 2017 and March 2022. After a cleaning process the dataset was split into training (67%), test (22%), and validation (11%) cohorts. A logistic regression model was trained to predict the pLoS in two classes: less than or greater than 7 days. The performance of the model was evaluated using accuracy, precision, sensitivity, specificity, and AUC-ROC metrics. The feature selection method returns nine interpretable variables, enhancing the models' transparency. In the validation cohort, the pLoS model achieved a specificity of 0.83 (95% CI, 0.82-0.84), sensitivity of 0.64 (95% CI, 0.62-0.65), accuracy of 0.76 (95% CI, 0.76-0.77), precision of 0.67 (95% CI, 0.66-0.69), and AUC-ROC of 0.82 (95% CI, 0.81-0.83). The model exhibits strong predictive performance and offers insights into the factors that influence prolonged hospital stays. This makes it a valuable tool for hospital management and for developing future intervention studies aimed at reducing pLoS.

Figures (6)

• Figure 1: Workflow of data preprocessing, feature selection, and model development. The figure shows the number of admissions at each stage of the pipeline, starting from the original hospital discharge database and after data cleaning and missing outcome handling. The dataset was restricted to features available at admission and reduced through a feature selection process based on Weight of Evidence (WoE) and Information Value (IV). The final cohort was split into training (67%), test (22%), and validation (11%) sets, and used for model training and evaluation.
• Figure 2: Receiver Operating Characteristic (ROC) curve for the test dataset, illustrating the model’s discriminative performance between short and prolonged hospital stays across different classification thresholds.
• Figure 3: Calibration curve for the test dataset, showing the agreement between predicted probabilities and observed outcomes, indicating a good calibration of the model.
• Figure 4: Confusion matrix for the test dataset, summarizing the classification performance of the model for short and prolonged hospital stays.
• Figure 5: Histogram of the logistic regression coefficients corresponding to the most important features selected by the model. The figure displays the magnitude and direction of the estimated weights, highlighting the reduced set of variables obtained through the feature selection process and supporting the interpretability of the final model.