Answers are grounded in this paper and cite their sources.
Patient generated text such as secure messages, surveys, and interviews contains rich expressions of the patient voice (PV), reflecting both communicative behaviors and social determinants of health (SDOH). Traditional qualitative coding methods are labor intensive and do not scale to the large number of patient-authored messages produced across health systems. Existing machine learning (ML) and natural language processing (NLP) approaches provide partial solutions for analyzing patient generated text. However, most methods treat patient generated text and SDOH as separate tasks, ignoring overlapping expressions within the same message. Some of them rely on pre-trained models that are not well suited to patient-facing language. These limitations highlight the need for scalable, domain-adapted approaches that jointly model the linguistic and social dimensions of patient generated communication. We introduce PVminer, a domain-adapted NLP tool for structuring patient voice in secure patient-provider communication. PVminer formalizes PV detection as a multi-label, multi-class prediction task and integrates three components: pre-training of patient-specific BERT encoders (PV-BERT-base and PV-BERT-large), unsupervised topic modeling for thematic augmentation (PV-Topic-BERT), and fine-tuned classifiers of Code, Subcode, and Combo-level labels. Topic information derived from a PV-Topic-BERT model is appended during both fine-tuning and inference to enrich input representations with high-level semantic cues. PVminer with our proposed domain-adapted pre-trained models achieves strongest performance across all three hierarchical tasks and outperforms biomedical and clinical pre-trained baselines. For Code-level classification, PV-BERT-large attains an F1 of 82.25%. At the Subcode level, PV-BERT-base achieves the highest F1 of 80.14%. For the more challenging Combo-level task, PV-BERT-base and PV-BERT-large obtain F1 of 77.58% and 77.87%. An ablation study further shows that author identity and topic-based augmentation each contribute meaningful gains. The training and evaluation code is available at https://github.com/Data-Mining-Lab-Yale/PVBERT .
No takes yet. Share an insight, caveat, or question.
Fodeh et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: