Predicting which digital health companies will succeed is challenging because these firms operate at the intersection of rapid software innovation and slow, regulation- and reimbursement-driven healthcare adoption. We developed a machine learning model, enhanced with large language model (LLM)-derived features from unstructured text, to predict digital health company outcomes and benchmarked it against expert investors. Using data on 10,245 companies founded since 2000 and observed through March 2025, we trained logistic regression, random forest, and XGBoost models, plus an ensemble, to predict successful exits. The ensemble achieved the best discrimination (AUC = 0.934 ± 0.017) and balanced performance (macro-F1 = 0.775 ± 0.031). Adding LLM-derived features improved AUC by 0.007 (P < 0.05). In a blinded benchmark, model accuracy (63%) exceeded mean investor accuracy (52%) and matched the top investor (62%). These results highlight the potential of hybrid human–AI approaches to enhance objectivity and efficiency in digital health investment decisions.
Pfitzer et al. (2026) studied this question.