Clinical Scorecard: Automated Detection of Fall-Associated Injuries in Unstructured Clinical Documentation

At a Glance

Key Highlights

• FRIs are a leading cause of emergency visits and hospitalizations among older adults, with significant healthcare costs.
• Manual review of electronic health records for FRIs is time-consuming and error-prone; NLP offers an efficient alternative.
• RoBERTa-based NLP model demonstrated high accuracy (precision 0.90, recall 0.91, F1 score 0.91, AUROC 0.96) in identifying FRIs from clinical notes.

Guideline-Based Recommendations

Diagnosis

• Use comprehensive FRI-related keyword lists to scan clinical notes for potential fall-related injury mentions.
• Apply validated NLP models, such as RoBERTa, to identify FRIs accurately in unstructured electronic health records.

Management

• Leverage automated detection of FRIs to enhance clinical research and potentially inform patient care strategies.
• Incorporate NLP-based identification tools to support timely recognition of fall-related injuries in older adults.

Monitoring & Follow-up

• Continuously validate and update NLP models with expert-labeled data to maintain high detection performance.
• Monitor model performance metrics including precision, recall, F1 score, AUROC, and AUPR to ensure reliability.

Risks

• Potential misclassification if keyword lists are incomplete or if clinical notes lack sufficient detail.
• Reliance on NLP models requires ongoing validation to mitigate errors due to unstructured data variability.

Patient & Prescribing Data

Medicare beneficiaries aged 65 years or older with clinical notes from multiple hospital settings

Automated identification of FRIs can improve research efficiency and potentially guide clinical interventions for fall prevention.

Clinical Best Practices

• Utilize large, diverse datasets from multiple hospital systems to train and validate NLP models for FRI detection.
• Engage clinical experts to label training data and validate NLP model outputs to ensure clinical relevance and accuracy.
• Implement multi-stage training of BERT-based models including masked language modeling and question-answering tailored to FRIs.
• Stratify patient cohorts to capture FRI instances across inpatient, outpatient, and other care settings for balanced model training.

References

• Mass General Brigham Health-Care System Data Source
• International Classification of Diseases, Tenth Revision (ICD-10)
• Natural Language Processing in Clinical Research