Sparse Mahalanobis Conformal Scoring was evaluated on approximately eight million outpatient visits and up to 12,829 diagnosis labels. It achieved Micro-F1 close to the strongest threshold comparator and the highest exact-match ratio among flexible-size rules, while retaining smaller regions and fewer labels than other nonconformity scores. This supports coding assistance, not unattended automated coding.
Key findings
- The primary output had Micro-F1 close to the strongest threshold comparator and the highest exact-match ratio among flexible-size rules. Mahalanobis scoring preserved point performance while retaining smaller regions with fewer distinct labels. Candidate-space adequacy remained central to coverage.
Why this matters globally
Large coding systems have long tails and consequential errors. Reviewable uncertainty sets could support safer billing and epidemiologic data, but conformal guarantees depend on calibration assumptions and distribution stability.
Thai researcher contribution
Researchers in Biomedical Sciences and Biomedical Engineering at Prince of Songkla University's Faculty of Medicine developed the framework using large-scale Thai outpatient data.
Limitations to consider
Data came from one tertiary hospital and labels may reflect historical workflow rather than clinical truth. The abstract does not report numeric coverage, set size, Macro-F1, or external validation. Coding-policy and distribution shifts may disrupt calibration, especially for rare labels.
Verify the original sources
Big Data and Cognitive ComputingRead the original article↗DOI: 10.3390/bdcc10070232