Work

Predicting operational risks in an O2O home nursing platform in China: machine learning models with SHAP-based explanations

Year: 2026

Type: article

Abstract: The rapid expansion of Online-to-Offline (O2O) home nursing services addresses the growing continuity of care demands of aging populations but introduces complex operational risks due to spatial dispersion and dynamic community environments. Traditional retrospective management approaches are inadequate for addressing these highly non-linear spatiotemporal uncertainties, highlighting an urgent need for proactive risk stratification. To develop, compare, and interpret predictive models utilizing traditional machine learning and Tabular Transformer architectures for stratifying two core operational risks (service overtime and long service duration) and one fee-structure classification task (high-cost orders) in digital nursing delivery. This retrospective observational study analyzed 89,623 continuous operational records from the “Yujian Nursing at Home” platform in 2025. The dataset was strictly partitioned into training, validation, and independent test sets. We compared the predictive performance of six traditional baseline classifiers against deep learning architectures, specifically comparing a simplified Minimal Transformer and a standard TabTransformer. Model efficacy was evaluated using the Area Under the Receiver Operating Characteristic Curve (AUC), Sensitivity, F1-score, and Precision-Recall curves. TreeSHAP for the Random Forest models and KernelSHAP for the TabTransformer were applied to extract global feature importance and interpret the decision-making mechanisms. In the full operational monitoring scenario, the Tabular Transformer achieved an AUC of 0.934 for high-cost order risks, while Random Forest achieved AUCs of 0.789 and 0.794 for long service duration and overtime risks, respectively. However, under the strict pre-dispatch setting (using only features available at order submission), predictive performance experienced an expected decrease: AUCs were 0.856 for high-cost risks (TabTransformer), 0.726 for long duration, and 0.744 for overtime (Random Forest). Identifying minority class cases in the highly imbalanced overtime task (9.78% positive rate) remained challenging under fixed thresholds, resulting in a low positive predictive value (PPV). SHAP analysis revealed that geospatial impedance and specialized service items were the features with the largest model-attributed contributions across all tasks. Advanced machine learning algorithms can stratify multidimensional operational risks in O2O home nursing platforms. By quantifying the impact of geospatial impedance and clinical complexity, these interpretable models facilitate a critical paradigm shift from retrospective quality control to proactive, algorithm-driven nursing governance, potentially generating hypotheses for optimizing human resource allocation, exploring interventions to mitigate occupational burnout, and informing equitable digital health policies.

Source: BMC Nursing

DOI: 10.1186/s12912-026-05269-5