This model performs Speaker Role Identification in Clinical Conversations between doctors, patients, medical assistants, and others. The model expects text sequences as input. These text sequences are expected to come from a speech transcript generated using ASR with speaker diarization. Inputs are formatted as follows:
bash
1Speaker 0: Do you have any pain today? (10)2Speaker 1: Yes, my chest has been hurting since yesterday. (11)
where the diarization identifiers have been pre-fixed to each utterance, the number of preceding utterances in the conversation post-fixed, and the target utterance is given as the last utterance.
Model Details
This model is an instance of the BertModelForSequenceClassification model available on HuggingFace, which has been fine-tuned for speaker role identification.
This model was trained on Fold 1 of our data, with two-utterance context windows, and 100% diarization accuracy, which we simply simulated from ground truth labels. Hence the name of the model:
This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.
Developed by: Andrew Zolensky, M.S., Member of AI4AI Lab and PhD student at @UPenn
Funded by [optional]: NIH, award number DP1-LM-014558
Citation [optional]: Zolensky A, Jang KJ, Sabin J, Hartzler A, Alasaly B, Mopidevi S, Liberman M, Johnson K. Speaker Role Identification in Clinical Conversations. medRxiv [Preprint]. 2025 Oct 11:2025.08.14.25332837. doi: 10.1101/2025.08.14.25332837. PMID: 41282782; PMCID: PMC12632672.
Model type: Transformer
Language(s) (NLP): English
License: []
Finetuned from model [optional]: BertForSequenceClassification
This model is intended for use by researchers interested in studying clinical communication profiles and their impact on functional outcomes at scale using NLP.
Out-of-Scope Use
This model will not work well for conversations where the roles of interest differ substantially from those of "patient," "provider", and "medical assistant", meaning that "other" will not be informative enough.
This model may not work well for data from non-clinical settings.
Training Details
Training Data
Training data are not made publicly available due to status as Protected Health Information.
Data come from the Establishing Focus Study, and RCT designed to study agenda-setting behavior in clinicians.
Training Procedure
Full fine-tuning using HuggingFace and BertForSequenceClassification library support.
