Enterprise AI Analysis
PVminer: A Domain-Specific Tool to Detect the Patient Voice in Patient Generated Data
PVminer is an NLP framework designed to extract and categorize 'patient voice' (PV) from patient-generated text, such as secure messages and survey responses. It integrates domain-adapted BERT pre-training, topic modeling, and multi-label classification to identify communicative behaviors and social determinants of health (SDoH) expressions across hierarchical levels (Code, Subcode, Combo). The framework outperforms general and biomedical baselines, demonstrating the value of in-domain adaptation and topic augmentation for robust PV detection.
Authors: Samah Fodeh, Linhai Ma, Yan Wang, Srivani Talakokkul, Ganesh Puthiaraju, Afshan Khan, Ashley Hagaman, Sarah Lowe, Aimee Roundtree
Key Impact Metrics for Your Enterprise
Leverage these insights to understand the tangible benefits of integrating PVminer into your healthcare communication analytics.
0
Code-Level F1 Score (PV-BERT-large)
0
Subcode-Level F1 Score (PV-BERT-base)
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
Patient-generated text offers rich insights into patient voice (PV), encompassing both communicative behaviors and social determinants of health (SDoH). Traditional qualitative coding methods are labor-intensive and unscalable for large datasets. Existing machine learning and NLP approaches are often limited, treating PCC and SDoH separately, relying on single-label classification, and using pre-trained models not adapted to patient-facing language. This leads to an incomplete understanding of the nuanced, overlapping expressions in patient communication. PVminer addresses these limitations by providing a scalable, domain-adapted framework that jointly models linguistic and social dimensions.
PVminer formalizes PV detection as a multi-label, multi-class prediction task. It integrates three core components: 1) Patient-specific BERT pre-training: Two models (PV-BERT-base and PV-BERT-large) are pre-trained on 6M unlabeled patient messages for domain adaptation. 2) Unsupervised topic modeling: A PV-Topic-BERT model, built using BERTopic on 500K unlabeled messages, generates latent thematic structures and keywords. 3) Fine-tuned classifiers: PV-BERT models are augmented with top keywords from PV-Topic-BERT and fine-tuned on an annotated dataset to predict Code, Subcode, and Combo labels. Author identity is also prepended as a special token to enrich input representations.
PVminer achieves strong performance across all hierarchical tasks, outperforming biomedical and clinical pre-trained baselines. For Code-level classification, PV-BERT-large attains an F1 score of 82.25%. At the Subcode level, PV-BERT-base achieves the highest F1 score of 80.14%. For the more challenging Combo-level task, PV-BERT-base and PV-BERT-large obtain F1 scores of 77.58% and 77.87%. An ablation study confirms that author identity and topic-based augmentation significantly contribute to performance gains. The framework's ability to handle multi-label, multi-class predictions reflects the layered nature of patient communication.
Impact of Domain Adaptation
82.25% F1 for Code-level classification with PV-BERT-large, outperforming general-domain BERT (80.97%) and BioBERT (78.64%). This highlights the critical importance of pre-training on patient-generated text.Enterprise Process Flow
Pre-train PV-BERT on 6M unlabeled messages
→
Train PV-Topic-BERT on 500K messages for themes
→
Augment PV-BERT with Topic Cues & Author Identity
→
Fine-tune for Multi-label PV Classification
→
Structured Patient Voice Representation
PVminer integrates domain-adapted pre-training, topic modeling, and fine-tuned classification for robust patient voice detection. This modular architecture allows for the flexible substitution of BERT encoders while leveraging large-scale unlabeled data for effective domain adaptation.
| Model | Code-Level F1 | Subcode-Level F1 | Combo-Level F1 |
|---|---|---|---|
| PV-BERT-large (PVminer) | 82.25% | 79.84% | 77.87% |
| BERT-large-uncased | 80.21% | 78.15% | 75.45% |
| BioBERT | 78.64% | 76.99% | 74.70% |
| SciBERT | 78.65% | 76.97% | 74.61% |
The comparison clearly shows PVminer's superior performance across all classification levels, especially with its domain-adapted PV-BERT models. This underscores the benefit of tailoring language models to patient-generated text over general or biomedical corpora.
Clinical Application: Triage & Support
PVminer can significantly enhance clinical workflows by automatically identifying critical patient needs and concerns. For instance, detecting 'Economic Stability' issues (an SDoH subcode) or 'PartnershipPatient_expressOpinions' (a communication subcode) can trigger timely referrals to support services or direct messages to care teams for prompt follow-up. This structured data enables better triage, reduces delays in managing side effects or treatment barriers, and improves overall patient engagement and outcomes.
Highlight: Automated recognition of these patterns may help care teams triage messages, prioritize high risk cases, and reduce delays in managing side effects or treatment barriers.
Calculate Your Potential ROI with AI
Estimate the time and cost savings your organization could achieve by automating patient voice analysis.
Annual Savings
$0
Hours Reclaimed Annually
0
Your AI Implementation Roadmap
A structured approach to integrating PVminer effectively into your operations, ensuring smooth adoption and maximum impact.
Phase 1: Data Preparation & Pre-training
Collection and de-identification of patient-generated text. Pre-training of PV-BERT and PV-Topic-BERT models on large unlabeled corpora.
Phase 2: Annotation & Model Fine-tuning
Expert annotation of a subset of patient messages with Code, Subcode, and Combo labels. Fine-tuning of PV-BERT models with topic augmentation and author identity.
Phase 3: Validation & Integration
Rigorous evaluation of PVminer's performance. Integration of the framework into existing clinical communication platforms for real-time patient voice detection.
Phase 4: Monitoring & Iterative Improvement
Continuous monitoring of model performance and data drift. Iterative refinement based on feedback and new data to enhance accuracy and coverage.
Ready to Transform Your Patient Insights?
Our experts are ready to discuss a tailored AI strategy that aligns with your enterprise goals. Book a free consultation today.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat