Prediction of long-term hospitalisation and all-cause mortality in patients with chronic heart failure on Dutch claims data: a machine learning approach

Abstract

Background Accurately predicting which patients with chronic heart failure (CHF) are particularly vulnerable for adverse outcomes is of crucial importance to support clinical decision making. The goal of the current study was to examine the predictive value on long term heart failure (HF) hospitalisation and all-cause mortality in CHF patients, by exploring and exploiting machine learning (ML) and traditional statistical techniques on a Dutch health insurance claims database. Methods Our study population consisted of 25,776 patients with a CHF diagnosis code between 2012 and 2014 and one year and three years follow-up HF hospitalisation (1446 and 3220 patients respectively) and all-cause mortality (2434 and 7882 patients respectively) were measured from 2015 to 2018. The area under the receiver operating characteristic (ROC) curve (AUC) was calculated after modelling the data using Logistic Regression, Random Forest, Elastic Net regression and Neural Networks. Results AUC rates ranged from 0.710 to 0.732 for 1-year HF hospitalisation, 0.705–0.733 for 3-years HF hospitalisation, 0.765–0.787 for 1-year mortality and 0.764–0.791 for 3-years mortality. Elastic Net performed best for all endpoints. Differences between techniques were small and only statistically significant between Elastic Net and Logistic Regression compared with Random Forest for 3-years HF hospitalisation. Conclusion In this study based on a health insurance claims database we found clear predictive value for predicting long-term HF hospitalisation and mortality of CHF patients by using ML techniques compared to traditional statistics.

The different risk of new‐onset, chronic, worsening, and advanced heart failure: A systematic review and meta‐regression analysis

“…Following the title and abstract screening process, 292 studies were deemed eligible for full‐text screening. Ultimately, during the final phase of full‐text screening, 65 studies (comprising 88 cohorts) met the inclusion criteria and were included in the subsequent meta‐analysis 10–74 . One additional study was identified through a manual cross‐reference of the included studies, bringing the total to 66 studies 75 (online supplementary Figure ).…”

Inter operator variability of machine learning researchers predicting all-cause mortality in patients admitted to intensive care unit

“…When building machine learning models for healthcare applications, algorithms need to be compared to appropriate baselines or standards to ensure the performance of these models is high enough to warrant the cost of deploying and maintaining them [26]. Performance of models is typically evaluated in relation to other outcomes [35], [36], [37], [38], [39], [40] or other prediction models [35], [41], [42], [43], [32], [44], [45]. To this end, we test the models built in this study which predict sudden death and other catastrophic cardiovascular events in two ways.…”

Use of big data from health insurance for assessment of cardiovascular outcomes

“…The analysis of insurance claims data for clinical findings steadily grew over the last years (Sawicki et al, 2020 ). Initially descriptive statistical methods were used, recently more advanced statistical and machine learning (ML) approaches have been employed in the clinical research (van der Galiën et al, 2021 ). Furthermore, there is a massive interest in the integration of multiple databases from different insurance providers (Riedel et al, 2018 ; Sabaté et al, 2021 ) or with the clinical data (Jones et al, 2018 ).…”

“…Depending on the country and type of health insurance, certain insurance populations might not accurately represent the general population of the country regarding social status, sex, and age (Barth et al, 2018 ; van der Galiën et al, 2021 ). Therefore, validation studies have been performed for different HIC datasets.…”

“…However, other ML-based techniques are scarcely applied (Rumsfeld et al, 2016 ). Using the state of art ML techniques on clinical data combined with HIC data might provide a deeper understanding of the patient's behavior, and improve risk stratification (van der Galiën et al, 2021 ). Despite the potential benefits of using ML models, caution needs to be applied; training datasets must be adequately defined, unbiased for sex, ethnicity, social and cultural biases.…”