Using machine learning to reduce unnecessary rehospitalization of cardiovascular patients in Saudi Arabia

Objectives: Patient readmission is a costly and preventable burden on healthcare systems. The main objective of this study was to develop a machine-learning classification model to identify cardiovascular patients with a high risk of readmission. Methods: Inpatient data were collected from 48 Ministry of Health hospitals (MOH) in Saudi Arabia from 2016 to 2019. Cardiovascular disease (CVD)-related diagnoses were defined as congestive heart failure (HF), ischemic heart disease (IHD), cardiac arrhythmias (CA), and valvular diseases (VD). Hospitalization days, daily hospitalization price, and the price of each basic and medical service provided were used to calculate the healthcare utilization cost. We employed a Python machine-learning model to identify all-cause 30-day CVD-related readmissions using the International Classification of Diseases, Revision 10 classification system (ICD10) as the gold standard. Demographics, comorbidities, and healthcare utilization were used as the independent variables. Results: From 2016 to 2019, we identified 403,032 hospitalized patients from 48 hospitals in 13 administrative regions of Saudi Arabia. Out of these patients, 17,461 had a history of hospital admission for cardiovascular reasons. The total direct cost of overall hospitalizations was 1.6 B international dollars (I$) with an average of I$ 3,156 per hospitalization, whereas CVD-related readmission costs were estimated to be I$ 14.9 M, with an average of I$ 7,600 per readmission. Finally, an empirical approach was followed to test several algorithms to identify patients at high risk of readmission. The comparison indicated that the decision-tree algorithm correctly classified 2,336 instances (926 readmitted and 1,410 not readmitted) and showed a higher F1 score than other models (64%), with a recall of 71% and precision of 57%. Conclusion: This study identified IHD as the most prevalent CVD, and hypertension and diabetes were found to be the most common comorbidities among hospitalized CVD patients. Compared to general encounters, readmission encounters were nearly two times higher on average among the study population. Furthermore, we concluded that a machine-learning model can be used to identify CVD patients at a high risk of readmission. Further research is required to develop more accurate models based on clinical notes and laboratory results.