The deep learning model detected 72.3% of heart failure rehospitalisations with 6.7% false positives, outperforming an XGBoost model by 4.6% in detection.
Does a deep learning model improve the detection of hospitalization risk in telemonitored heart failure patients compared to a baseline machine learning model?
A transformer-based deep learning model analyzing daily vital parameters outperforms traditional machine learning in predicting heart failure rehospitalizations.
Absolute Event Rate: 0% vs 0%
Abstract Background Heart Failure (HF) represents end-stage myocardial disease, expected to increase with an aging population. HF accounts for a significant proportion of hospitalisations, and due to the unpredictable nature of the syndrome which often involves multiple comorbidities, can lead to frequent readmissions post discharge. Remote patient management through telemonitoring has shown to be a promising approach for early detection of worsening health status through continuous logging of vital parameters (VP). Implemented systems today often use predefined thresholds to determine instances of high risk, ignoring sequential- and temporal dependencies in the data. More sophisticated model architectures have the potential to use these data sources to solve tasks, where deep learning (DL) often outperforms more traditional machine learning (ML) architectures by using larger and more complex networks. Purpose Create a DL system for detecting the deterioration of the patient's condition by estimating the probability of hospitalization in telemonitored HF patients. Methods A real-world cohort of out-patients with HF undergoing telemonitoring for new or worsening HF, with daily measurements of VP (blood pressure (BP), heart rate (HR) and weight) during 6 months, were analysed. The data sequences for each VP were filtered with clinically relevant thresholds to remove outliers and subsequently divided into shorter time windows (length 10-14 measurements) with overlapping boundaries. Temporal restrictions to find measurements for a window were twenty-one days. The data windows were separated into training- and validation sets based on the patient origins, ensuring model validation on unseen information. We constructed the DL model using three transformer-based encoders, one for each of the two BP variables and the HR, followed by task-specific secondary networks. Frozen linear networks were used as projection heads for the anomaly-based contrastive pre-training and trainable, fully-connected linear layers were used as the classifiers (Figure 1). Results We analysed n=278 patients (n=18 rehospitalisations related to HF during the monitoring period, mean age of 67.6 years, n=72 women (25.9%) 227 patients (81.7%) were classified with HFrEF). The imbalance ratio for the dataset was 222.2. The DL model was able to correctly detect 72.3% of the rehospitalisations whilst overestimating 6.7% of non-rehospitalisations (Table 1). In comparison to a baseline ML model, the DL model outperformed the tested XGBoost model on both sensitivity and F1-score for the unseen validation data. Furthermore, the anomaly-based contrastive pre-training increased the ability to detect hospitalisation with 4.6%. Conclusions The DL model resulted in a higher performance on all evaluation metrics compared to the tested ML model. The model shows promise in identifying adverse outcomes within a time frame that allows for counteractive treatment in patients with HF.Graphical illustration of the DL model
Aerts et al. (Sat,) reported a other. The deep learning model detected 72.3% of heart failure rehospitalisations with 6.7% false positives, outperforming an XGBoost model by 4.6% in detection.