Introduction
In modern healthcare systems, the Diagnosis Related Groups (DRG) payment model has become a standard for hospital reimbursement. Under DRG, payments are determined by a patient's primary diagnosis and additional diagnoses, such as complications (CC) or serious complications (MCC). A significant challenge in this model is the "write-missing" diagnosis—a condition identified by a physician in clinical notes but omitted from the final discharge summary. Unlike a "missed diagnosis," where a condition is never spotted, a write-missing diagnosis represents a clerical or cognitive oversight that leads to incomplete medical records and substantial financial losses for hospitals.
Researchers from Tsinghua University have addressed this by developing a deep learning pipeline designed to act as an automated "safety net," identifying these omitted entries in Chinese Electronic Medical Records (EMRs). By automating the audit process, the system aims to improve both medical record accuracy and hospital financial health.
Methodology
The proposed system utilizes a three-stage pipeline to systematically identify and verify missing diagnoses.
1. Disease Recall Module
The first stage involves extracting all potential disease mentions from the full EMR text. This is achieved through a combination of extensive medical vocabularies (approximately 40,000 terms) and Named Entity Recognition (NER) techniques to pull every possible candidate from the patient's history and ward round records.
2. Disease Context Logic Judgment
This core module determines if a mentioned disease is a confirmed diagnosis for the current hospital visit. It must distinguish between active conditions and mentions of "family history," "ruled out" symptoms, or "history of" previous illnesses. The researchers implemented a BERT-based encoder enhanced with a gating mechanism to incorporate domain-specific features.
Key innovations in this module include:
- Feature Integration: The model incorporates disease location, negative symptom indicators (Neg embedding), and serial number information (order embedding) to recognize when doctors list diagnoses in numbered sequences.
- Data Augmentation: To overcome the scarcity of labeled medical data, techniques like Easy Data Augmentation (EDA) and "Disease Replacement" were used to generate diverse training examples from a limited set of 500 manually labeled records.
3. Disease Relationship Comparison
Finally, the system compares confirmed diagnoses against the existing discharge list. Because physicians often use synonyms, the team employed SimCSE (Simple Contrastive Learning of Sentence Embeddings) and BERT to evaluate semantic similarity, ensuring that a diagnosis isn't flagged as "missing" simply because of a variation in terminology.
Results and Impact
The system was validated using a dataset of approximately 10,000 EMRs from a tertiary hospital in China. The pipeline demonstrated strong performance across its modules, achieving a Precision of 0.925 and an F1-score of 0.807. Ablation studies highlighted that the "Disease Location" feature was the most critical signal for determining the validity of a diagnosis.
In a real-world test of 9,385 EMRs, the system identified 5,116 cases with missed diagnoses. This automation led to the adjustment of DRG groupings for 366 cases, recovering approximately 1,241,030 RMB (approx. $170,000 USD) in a single month. This represents a recovery of 7.49% of potential financial losses for the hospital under the DRG model.
Conclusion
This research demonstrates that deep learning can effectively bridge the gap between clinical documentation and formal medical billing. By focusing on "write-missing" diagnoses, the system provides an interpretable and financially viable solution for hospital administration. Future work aims to expand the system's capabilities to include "intelligent inference," allowing it to suggest diagnoses based on laboratory results and medication records even when not explicitly stated in the text.
Summary updated