PREDICTING HOSPITAL LENGTH OF STAY AMONG DIALYSIS PATIENTS USING EXPLAINABLE MACHINE LEARNING

Length of Stay (LOS) in the hospital among dialysis patients is perhaps the most critical indicator of the burden of care. While each additional day of stay increases the risk of complications, emotional distress, and financial cost, understanding the factors influencing outcomes during the first 10 days of admission may prove crucial. Unfortunately, classical models are limited in handling complex, high-dimensional information from personal and socioeconomic profiles, as well as clinical, biochemical, and dialysis-related predictors. This study examines the use of explainable artificial intelligence (XAI) to bridge this gap.

This retrospective study analysed 1,057 medical records from three tertiary hospitals in Malaysia. Using LOS as the target variable, XGBoost regression is trained to leverage discriminating information from 19 quantitative (e.g., serum creatinine, haemoglobin, phosphate, calcium, blood pressure, albumin, and dialysis frequency) and 50 qualitative predictors (e.g., demographic, comorbidity, and dialysis-related characteristics). The risk of overfitting is addressed through optimisation via a hybrid ensemble of five-fold cross-validation and comparing with ridge regression. Through an established 70:30 random split, model performance is evaluated using RMSE, MAE, and R². Feature importance with SHAP summary plots helps with model explainability.

The mean age of patients was 56.2 ± 12.8 years, with the majority being males (50.8%), on haemodialysis (84.1%), and primarily admitted due to fluid overload (20.3%). Commonest comorbidities were hypertension (86.4%), diabetes mellitus (73.2%), and anaemia (65.4%). The mean LOS was 4.9 ± 3.2 days. The optimised ensemble model achieved good accuracy during training (RMSE = 2.189, MAE = 1.842, and R2 = 0.539), while test-set performance supports its fit and generalizability (RMSE = 2.688, MAE = 2.250, and R2 = 0.271). The primary admission diagnosis, albumin, white blood cell count, serum calcium, potassium, platelet count, and systolic blood pressure emerged as the top predictors of LOS. For the primary admission diagnosis, SHAP plots indicate that catheter-related bloodstream infection (CRBSI) is the most important predictor of longer LOS. In contrast, those admitted for fluid overload and those electively admitted for vascular access were more likely to have shorter hospital stays. The plots also showed that normal albumin and platelet levels were protective against longer LOS, though the non-linear relationship visualisation showed no gain from higher-than-normal values.

This study demonstrates the feasibility of using an explainable ensemble machine learning approach for early LOS prediction and understanding, based on information commonly available at admission. CRBSI was the most predictive of longer LOS, while fluid overload, despite being associated with shorter LOS, was the commonest cause of admission. Assessing the burden of disease is inherently complex, but focusing on several key biochemical parameters on admission may help devise an effective treatment plan to reduce LOS. This study illustrates the use of an explainable AI model to achieve a novel understanding to assist with complex decision-making in the management of dialysis patients during hospital admission.