An Interactive Framework for Generating Clinical Data with Human Feedback

Evaluating machine-learning models in critical-care settings is particularly challenging. The lack of sufficient data for rare but clinically important cases can lead to unreliable model performance. Clinicians often require specific patient scenarios to assess the robustness of machine learning methods. However, it is difficult to manually construct such patient profiles. Generating patient data for some conditions presents a promising alternative.Therefore, conditional generation methods are needed to create realistic synthetic data that aligns with clinician-defined criteria.To address this challenge, we introduce a novel interactive generative framework that allows clinicians to specify desired patient characteristics and generate synthetic data accordingly.In this paper, we focus on the problem of generating synthetic data for electronic health records (EHR), especially for patients on mechanical ventilation and ECMO where the data is limited.We propose a novel interactive tool InterGenEHR that leverages the generative model with arbitrary conditioning to generate synthetic data conditioned on clinician-specified features.We evaluate our proposed interactive framework using numerical metrics of synthetic data quality and clinically meaningful assessments based on clinician feedback.We also provide a web application that allows clinicians to interactively generate synthetic data based on their requirements and evaluate via clinicians.In summary, we provide an effective tool for validating machine learning methods using clinician feedback tailored to individual patient scenarios.