Advancing efficacy prediction for electronic health records based emulated trials in repurposing heart failure therapies

Figures

• 

  Fig. 1. An overview of the proposed framework. a illustrates the general pipeline, starting with a search for relevant clinical trial drugs on ClinicalTrials.gov and identifying genes using the DisGeNET48 database. With a list of repurposed trial drugs and associated genes, the proposed drug-target prediction model predicts the probabilities of drug-gene associations. We evaluate the efficacy direction of repurposed drugs—beneficial (positive impact) or not—by integrating drug-target prediction with an EHR-based emulation protocol that compares exposed and unexposed groups using HF-related clinical tests. In this framework, the odds ratio, derived from the logistic regression coefficient, serves as a predictive component and is combined with predicted targets through a normalization transformation. Importantly, while the odds ratio is used to estimate efficacy direction in the ET-based comparison method, our proposed framework normalizes the odds ratio and integrates it with predicted drug targets to generate a probability-like score, representing the drug’s potential benefit in clinical trials (i.e., efficacy direction) compared to the standard of care. b provides details about the…

• 

  Fig. 2. The flow of the data generation for the study. A detailed data processing flow is presented to show the corresponding data size and the number of patients resulting from each processing step.

• 

  Fig. 3. Evaluation of the proposed drug-target prediction model using the BETA framework and predictions for 17 repurposed RCT drugs. a illustrates the model’s overall performance across all seven tests within the BETA framework, demonstrating that the proposed method significantly outperforms all baseline methods. b showcases the prediction results for 17 repurposed RCT drugs based on the proposed model. The AUC-ROC and PR-AUC curves are provided to demonstrate the model’s predictive performance. Note: * indicates statistically significant improvement tested with a paired one-tailed t-test, where * represents P-value < 0.05, ** for P-value < 0.01, *** for P-value < 0.001, and **** for P-value < 0.0001.

• 

  Fig. 4. Evaluation of EHR emulation with and without the integration of drug-target prediction. a displays all analyses conducted using the BASELINE cohort across 11 prognostic outcome measures (NT-proBNP, BNP, TnT, TnI, Creatinine, BUN, eGFR, Hemoglobin, Renin, Aldosterone, CRP). Two subplots are included: (i) EHR-based emulation and (ii) EHR emulation with integrated drug-target prediction. Detailed predictions for each of the 17 RCT drugs are shown, along with the overall AUC-ROC and PR-AUC for all drugs. Additionally, b presents analyses based on BNP, the best-performing prognostic outcome measure, across 4 HF patient cohorts (BASELINE, ALL, HFpEF, HFrEF), with two subplots: (i) EHR-based emulation and (ii) EHR emulation with integrated drug-target prediction. Please note, for (i) EHR-based emulation, while the odds ratio derived from the logistic regression coefficient can be used to estimate both the magnitude and direction of treatment effects, we focus solely on estimating the direction to align with the objectives of this study. Similarly, (ii) For EHR emulation with integrated drug-target prediction, our model predicts the likelihood of a drug being beneficial compared to the…

• 

  Fig. 5. Comparison of AUCROC between the EHR-based emulation and the ET integrated with drug-target prediction. a displays the comparison for each prognostic outcome measure, with a focus on the best-performing HF cohort, where the EHR-based emulation is highlighted in red, and the ET integrated with drug-target prediction is highlighted in blue. b shows the comparison based on (i) outcome measures and (ii) HF cohorts, respectively. As shown in (b), a significant improvement was observed in the ET integrated with drug-target prediction. Note: * indicates statistically significant improvement tested with a paired one-tailed t-test, where * represents P-value < 0.05, ** for P-value < 0.01, *** for P-value < 0.001, and **** for P-value < 0.0001.

• 

  Fig. 6. Results of incorporating interactions between HF drugs and gender in the proposed method (ET integrated with Drug-Target Prediction) for the top four prognostic markers that yielded the highest number of improved predictions. a Illustrates the performance of predictions based on various interaction combinations for the top four prognostic markers: (i) CRP, (ii) Creatinine, (iii) TnI, and (iv) BUN. Red nodes indicate an improvement in AUCROC, while blue nodes represent a decrease in AUCROC performance. The size of each node corresponds to the AUCROC score. b Provides a detailed view of the improved predictions for 17 drugs across the four prognostic markers.