Natural language processing improves the prediction of mortality after heart failure hospitalization
O El Khettari, V Barthet, G Hocquet, J Weller, L Pellet, M J Aroulanda, R Cador, A Buronfosse, P De Groote, M Komajda, P Zweigenbaum, E MorinAbstract
Aims
Early post-discharge mortality in heart failure (HF) remains high, yet most prediction tools rely on structured electronic health record (EHR) data and overlook prognostic information provided by clinical notes. We evaluated whether combining structured data with natural language processing (NLP) representations from clinical notes improves 90-day mortality prediction and enables risk stratification.
Methods
We analysed 2257 HF admissions (248 deaths at 90 days) from a single French centre (2015–2020). Deaths were identified from the National mortality registry. For each admission, we used: (i) structured variables available in electronic health records at discharge (demographics, comorbidities, laboratory, treatments)), and (ii) in-hospital textual clinical notes, processed with transformer-based models[1] to extract medically relevant concepts (symptoms, treatments, examinations, autonomy, behaviour) and converted into quantitative features. We evaluated three predictive strategies: structured-only (logistic regression), text-only (NLP transformers), and a hybrid model combining both. Performance was assessed with 5-fold cross-validation and reported as area under the ROC curve (AUC) with 95% confidence intervals (CI).
Results
Structured data alone showed a better discrimination (AUC 0.76, 95% CI 0.71–0.80) than the text-only model (AUC 0.71, 95% CI 0.68-0.74). The hybrid model integrating both structured and NLP-derived text features achieved the highest discrimination (AUC 0.79, 95% CI 0.76–0.83). Patients were stratified into low-, medium-, and high-risk groups based on the probabilities predicted by the hybrid method. Kaplan–Meier curves demonstrated a clear separation between risk categories, (with high-risk patients showing early and steep survival decline). Observed 90-day mortality was 2.3%, 3.9%, and 19.6% in patients classified as low-, intermediate-, and high-risk based on probability tertiles predicted by the hybrid method.
Conclusions
Combining structured EHR data with NLP-extracted information from clinical notes improves 90-day mortality prediction in HF compared with either source alone. This hybrid approach captures complementary prognostic details and enables robust stratification into low-, medium-, and high-risk groups, supporting intensive post-discharge management.For image description, please refer to the figure legend and surrounding text.