Document Type

Abstract

Publication Date

2026

DOI

10.1016/j.rpth.2026.106398

Publication Title

Research and Practice in Thrombosis and Haemostasis

Volume

10

Issue

Suppl. 2

Pages

1004-1005

Conference Name

2026 Congress of the International Society of Thrombosis and Haemostasis, July 11-15, 2026, Paris, France

Abstract

Background

Natural language processing (NLP) has been used successfully to detect pulmonary embolism (PE) in electronic health records (EHR). Federated Learning (FL) is a technique which allows two or more sites to collaborate and train a single model using individual datasets without the need to share any data. In the healthcare setting, this approach could potentially allow training models with improved accuracy and generalizability without compromising on privacy or security.

Aims

Assess the feasibility of using FL for training NLP transformer models for the detection of PE events on EHR corpora from separate institutions and evaluate the performance of this approach compared to the current single-centre training paradigm.

Methods

Training, validation, development and testing sets were derived in a 7:1:1:1 ratio for each dataset. The two datasets used were derived from the internal MSK-IMPACT cohort and the external MIMIC-IV-Ext-PE cohort. As previously described for the CEDARS NLP application, EHR documents were labelled as occurring before or after a PE event. The details and metadata for each dataset can be found in Table 1. Three fine-tuned NLP transformer models were trained, one on each individual dataset and one aggregated model using FL. The classification threshold for each model was set to target a sensitivity of 95% or higher on the development set and the final metrics were computed on the testing set of each cohort.

Results

See Table 2 for the results in the MSK and MIMIC datasets respectively. Using the aggregated model, positive predictive value and sensitivity were 0.644 (95% CI=0.529-0.754) and 0.920 (95% CI=0.835-0.990) respectively in the MSK dataset, compared to 0.615 (95% CI=0.530-0.698) and 0.970 (95% CI=0.931-1.000) in the MIMIC dataset. The PPV and specificity of the aggregated model were significantly superior (p< 0.001) to the individual models while maintaining statistically similar sensitivity and AUROC (area under the receiver operating characteristic curve).

Conclusions

Using FL, the aggregated model performed well to detect PE events in both individual cohorts. FL for NLP should be explored further in the healthcare setting, where it has the potential to improve model generalizability whilst safeguarding protected health information.

Original Publication Citation

Irani, K., Patell, R., lam, b., Kovalenko, I., Zwicker, J., & Mantha, S. (2026). Federated learning-based natural language processing for the detection of pulmonary embolism in electronic health records. Research and Practice in Thrombosis and Haemostasis. 10(Suppl. 2), 1004-1005. https://doi.org/10.1016/j.rpth.2026.106398

Download

Share

COinS