Discriminating value of artificial intelligence based models for heart failure readmissions and mortality: A comparison of patients included or not in the PRADO program

Background The PRADO program was designed to improve post-discharge patient care in heart failure (HF). In this context, prediction of readmission or death in HF is of critical importance and is currently based on a limited number of variables selected by experts/literature. Artificial intelligence (AI) allows to include a non-limited and non-selected number of variables. Purpose The objective was to predict the probability of HF readmission and/or death inpatients with a new diagnosis of HF included or not in the PRADO program at the cardiologist's discretion, using medical records and machine learning models, without a priori. Methods This pilot monocentric study included all patients having a first HF admission between January 1st, 2015 and December 31, 2018, enrolled by data record (ICD-10) from the department of cardiology Paris Saint-Joseph Hospital. One thousand variables extracted from electronic health records/local hospital discharge database (PMSI) were used to create models. Data from the national PMSI were also included using the Hawkes process. Models were constructed on patients admitted between 2015 and 2017 to predict 1 and 3 months probability of HF readmission and HF readmission or mortality. Discrimination was tested using the area under the ROC curves (AUC) inpatients admitted in 2018. Results The discrimination value was limited overall ( Table 1 ). It was better for the 1 month than for the 3 months prediction of clinical outcomes. The models predicted better the composite outcome than HF readmissions and performed better for non-PRADO patients. Conclusion The predictive value of these models using AI in our population is limited and performed better at 1 month than 3 months. The better discriminating value in non-PRADO patients might result from different clinical profiles, from the arbitrary nature of the selection by cardiologists and from the limited sample size compared to the high number of variables included in the models.