Interpretable Machine Learning-Based Clinical Predictive Model for Early Readmission in Patients with Cardiogenic Shock

Background/PurposeCardiogenic shock (CS) is a critical condition characterized by low cardiac output leading to end-organ hypoperfusion and often multisystem organ failure, affecting up to 50,000 people annually in the United States. Acute myocardial infarction (AMI) is the primary cause, responsible for 81% of CS cases. Despite advancements in reperfusion therapies improving survival, in-hospital mortality remains high at 40%-67%, with 18.6% of survivors readmitted within 30 days. Traditional methods struggle to quantify and process the complex interactions among various risk factors, making prediction of readmissions challenging. Machine learning (ML) offers a promising solution by capturing intricate patterns and non-linear relationships among numerous variables. This study aims to develop an ML-based prediction model for 7-day and 30-day readmission rates in CS patients using the 2019 National Readmission Database (NRD). Additionally, the study utilizes SHapley Additive exPlanations (SHAP) to interpret the outcomes of the applied machine learning methods. MethodWe conducted a retrospective study using the NRD for 2019. Index hospitalizations were identified by non-elective admissions with a primary ICD-10 diagnosis of cardiogenic shock. Exclusions included patients under 18, missing length of stay or days to event data, and same-day transfers. The primary outcome was readmission within 7- and 30-days post-discharge. Welchs t-test compared continuous variables. Various ML models were evaluated for their predictive performance, and SHAP values were used to interpret the most influential features. ResultsThe study included 97,653 adults hospitalized for CS, with a mean age of 65.8 years and 38.4% being female. The in-hospital mortality rate was 33.7%. Among 51,976 index hospitalizations, 8.3% were readmitted within 7 days, and 21.02% within 30 days. Significant predictors of higher readmission rates included younger age, lower income, Medicaid insurance, CKD3, drug abuse, chronic pulmonary disease, PHTN, depression, leukemia, lymphoma, discharge against medical advice, and certain hospital characteristics. The FT-Transformer (a specialized deep neural network approach for tabular data) model achieved the highest AUCs of 0.76 and 0.78 for 7-day and 30-day readmissions, respectively, outperforming traditional methods like Logistic Regression (AUCs: 0.60 and 0.63). SHAP analysis revealed a wide array of features contributing to readmission predictions at both the population and individual levels. For the general population, the top features included APRDRG, DRG_NoPOA, age, chronic kidney disease, length of stay, number of ICD-10 codes, and disposition at discharge. In contrast, for an individual patient, the most influential feature for predicting 7-day readmission may differ, though there are some overlaps. This highlights the potential of personalized medicine, where individual risk factors are weighted differently compared to the general population, providing tailored insights for targeted interventions. ConclusionThis study demonstrates that advanced ML models, particularly the FT-Transformer and Random Forest, significantly outperform traditional methods in predicting readmissions in CS patients. The use of SHAP values enhances the interpretability of these models, providing actionable insights for healthcare providers. The differentiation between general population feature contributions and individual-specific factors underscores the importance of personalized medicine. By understanding individual risk profiles, healthcare providers can implement more precise and effective interventions, ultimately aiming to reduce readmissions and optimize healthcare outcomes.