Artificial intelligence-based analysis of payment system data can predict one-year mortality of hospitalized patients with heart failure

Background Individual risk stratification is fundamental in the care of heart failure (HF) patients. However, the prediction performance of risk scores, such as MAGGIC and SHFM, is not adequate and, more importantly, they need additional predictors including various biomarkers, imaging data, and environmental factors. Data from a case-mix payment system including diagnosis and procedures with outcomes can be used to develop the risk prediction models, allowing the use of big data for a more accurate prediction of mortality. Purpose This study aimed to develop artificial intelligence (AI) models for predicting 1-year mortality in patients hospitalized due to HF. Methods We analyzed the data from 10175 patients enrolled in the Japanese Registry Of Acute Decompensated Heart Failure (JROADHF). Candidate variables included the data obtained from a payment system introduced by the Japanese government, the Diagnosis Procedure Combination (DPC), which included each patient profile (age, sex, height, weight), principal diagnosis for hospitalization, comorbidities, procedures, length of hospital stay, and discharge status. They did not include clinical data available from patients such as vital status, laboratory data including bio-makers, electrocardiographic and echocardiographic data. The collected data were divided into the training set and the validation set (80%: 20%). With the training set, 5 AI models (logistic regression, random forest, support vector machine, neural network, and ensemble classifier) learned the one-year mortality results. AI models were evaluated by using the validation set with ROC analysis. The training and validation steps were repeated 10 times with different seed values to calculate the C-statistic of each model. We also identified the predictors for one-year prognosis acquired from the AI models. Results At 1-year of follow-up, a total of 1727 patients had died (17%). Among the machine learning models, the ensemble classifier showed the highest C-statistic of 0.76 (95% confidence interval: 0.75 to 0.77) for predicting mortality. Top predictors acquired from the random forest classifier was ADL (Barthel Index) at discharge, age, body mass index, and length of hospital stay. Conclusion By using AI-based analysis of a national case-mix payment system data, the present risk stratification model could predict the one-year mortality of hospitalized HF patients without any quantitative laboratory and physiological data. Furthermore, the present results could emphasize the advantage of this approach using the claim-based data that are routinely collected in a usual daily practice with no need to collect any additional information. Funding Acknowledgement Type of funding source: Public grant(s) – National budget only. Main funding source(s): Japan Agency for Medical Research and Development