Article Sidebar
Main Article Content
Abstract
Background. Psychiatric risk changes during hospitalization. Document-level natural language processing can retain historical high-risk language after the current clinical state has changed.
Objective. To evaluate an explainable approach that reconciles chronology, source authority, negation, and transition barriers across longitudinal records against a maximum-risk baseline.
Methods. This retrospective study used all 180 synthetic inpatient psychiatric chart packets of ClinAuthBench v1. A deterministic pipeline gave priority to the most recent suicide-risk screens, psychiatric progress notes, and social-work transition-barrier notes. The baseline was a maximum-risk-across-record strategy. The primary endpoint was current risk (low versus elevated/uncertain). Secondary endpoints were three-class risk, diagnosis, psychosis, substance use, transition barriers, and lower-level-of-care readiness. We used accuracy, Wilson 95% confidence intervals (CIs), macro-F1, Cohen's κ, paired bootstrap differences, and exact McNemar testing.
Results. Gold current risk comprised 72 low, 82 moderate, and 26 high cases. Temporal reconciliation classified 180/180 cases correctly on the binary endpoint, versus 108/180 for the baseline (difference 40.0 percentage points; bootstrap 95% CI 32.8–47.2; exact McNemar p < 0.001). Three-class accuracy was 88.9% (160/180; 95% CI 83.5–92.7), macro-F1 was 0.852, and κ was 0.824, with all errors at the moderate–high boundary. Five structured phenotype endpoints reached 100% accuracy.
Conclusion. In this synthetic benchmark, temporal and source-aware reconciliation kept historical high-risk language from dominating current-risk classification. The results show construct-aligned technical validity. They do not show clinical deployment validity, and external validation on real notes with clinician adjudication is required. For every packet the explainable trace retained source, time, screen status, barrier status, and context.
Keywords
Article Details
As our aim is to disseminate original research article, hence the publishing right is a necessary one. The publishing right is needed in order to reach the agreement between the author and publisher. As the journal is fully open access, the authors will sign an exclusive license agreement.
The authors have the right to:
- Share their article in the same ways permitted to third parties under the relevant user license.
- Retain copyright, patent, trademark and other intellectual property rights including research data.
- Proper attribution and credit for the published work.
For the open access article, the publisher is granted to the following right.
- The non-exclusive right to publish the article and grant right to others.
- For the published article, the publisher applied for the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.
