Abstract 156: Insights From Meta-analysis Of Studies Using Machine Learning To Predict Mortality, Readmission, Or Other Outcomes Among Heart Failure Patients

Developing and implementing analytical models for predicting mortality or readmission related outcomes among heart failure (HF) patients are challenging. In this study, we used meta-analyses of reported predictive models to assess what machine learning (ML) has been able to accomplish in this field, by evaluating the ML model performance for studies in HF.

We performed a literature search using Google Scholar, Web of Science and PubMed. The studies reporting AUC and 95% CI for various models were included. In addition, total participants, year of publication, type of analytical method (logistic regression, RF, etc.) and type of outcome (mortality, readmission, etc.) were extracted. We combined effect sizes using random effects (RF) model, and tested for heterogeneity, and publication bias.

12 studies were included in the analysis (patients= 123,832; AUC=15, with outcome mortality =17,471, readmission=15,703, hospitalization=67,523). Combined mean AUC was 0.77 (95% CI: 0.72, 0.82). Test of heterogeneity showed high variation between studies (I2=98.9%). Egger’s test intercept was 5.2 (95% CI: -4.2, 14.7, p > .25) indicating no small study effects/bias. Meta regression showed newer publications provide better AUC values (p < 0.03). In subgroup analysis, the pooled AUC for readmission, hospitalization, and mortality groups were 0.71, 0.80, and 0.78 respectively. The highest individual AUC was from neural networks (NN) predicting hospitalization with AUC 0.96 and lowest was from RF predicting readmission with AUC 0.65.

Presented models were diverse, ranking in quality from fair to very good, and being varied for different clinical outcomes among HF patients. Situation known from studies using classical statistical methods holds also for methods using ML, with better predictive values for hospitalizations, and lower for other outcomes. Methods using NN and methods using higher numbers of variables performed very well and had highest predictive power.