Development and Internal Technical Evaluation of a Transparent Dashboard for Indonesian Medical Malpractice Disputes: A Pilot Study

Abstract

Background. Judicial assessment of suspected medical malpractice requires simultaneous consideration of clinical facts, documentation, professional standards, causation, organizational responsibility, and procedural proof.

Objective. To develop and technically evaluate a transparent local dashboard that structures these elements for education and research without substituting for expert or judicial judgment.

Methods. Sixteen de-identified Indonesian disputes (2006–2026) were purposively assembled, one record per dispute. Eight prespecified concepts were manually coded and compared with a regular-expression dictionary, free of outcome terms, applied to evidence notes (dictionary-based natural language processing [NLP]). Five features entered a ridge logistic model (λ = 1) evaluated by leave-one-dispute-out cross-validation. Six families of deterministic software tests were run.

Results. The corpus contained six granted and ten rejected outcomes. Across 128 concept decisions, the dictionary produced 33 true positives, 2 false positives, and 13 false negatives (micro-precision 94.3%, micro-recall 71.7%, micro-F1 81.5%; macro-F1 87.9%). Severe injury was positive in 15 of 16 disputes and had the lowest recall (0.40). Out-of-fold sensitivity was 5/6 and specificity 10/10; the Brier score was 0.050 with manual features and 0.061 with NLP-derived features. All 49 software tests passed.

Conclusion. The prototype provides traceable exploration of a small medicolegal corpus. Because one abstractor produced the labels and notes, and the predictors restate the courts’ own grounds, the near-perfect internal separation is not evidence of prediction or external validity. Use should be limited to education, hypothesis generation, and abstraction quality assurance.