{"paper":{"title":"Bridging the Domain Divide: Supervised vs. Zero-Shot Clinical Section Segmentation from MIMIC-III to Obstetrics","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Supervised clinical section segmentation models drop in performance when moving from MIMIC-III to obstetrics notes, while zero-shot models remain robust after correcting for hallucinated headers.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Barbara Di Eugenio, Baris Karacan, Patrick Thornton","submitted_at":"2026-02-19T16:25:07Z","abstract_excerpt":"Clinical free-text notes contain vital patient information. They are structured into labelled sections; recognizing these sections has been shown to support clinical decision-making and downstream NLP tasks. In this paper, we advance clinical section segmentation through three key contributions. First, we curate a new de-identified, section-labeled obstetrics notes dataset, to supplement the medical domains covered in public corpora such as MIMIC-III, on which most existing segmentation approaches are trained. Second, we systematically evaluate transformer-based supervised models for section s"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"while supervised models perform strongly in-domain, their performance drops substantially out-of-domain. In contrast, zero-shot models demonstrate robust out-of-domain adaptability once hallucinated section headers are corrected.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The new obstetrics dataset is representative of the broader domain and that manual correction of hallucinations provides a fair, scalable basis for comparing model performance.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Supervised clinical section segmentation models perform strongly in-domain on MIMIC-III but degrade substantially out-of-domain on a new obstetrics dataset, whereas zero-shot LLMs show robust cross-domain performance after hallucination correction.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Supervised clinical section segmentation models drop in performance when moving from MIMIC-III to obstetrics notes, while zero-shot models remain robust after correcting for hallucinated headers.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"f84110bdb8561e7df3e0ccbe2c5f9530ae2e610cdccf4a587f35660397e0db1d"},"source":{"id":"2602.17513","kind":"arxiv","version":2},"verdict":{"id":"606d2148-e22b-4e64-bf15-5e66a636719a","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T20:51:55.012614Z","strongest_claim":"while supervised models perform strongly in-domain, their performance drops substantially out-of-domain. In contrast, zero-shot models demonstrate robust out-of-domain adaptability once hallucinated section headers are corrected.","one_line_summary":"Supervised clinical section segmentation models perform strongly in-domain on MIMIC-III but degrade substantially out-of-domain on a new obstetrics dataset, whereas zero-shot LLMs show robust cross-domain performance after hallucination correction.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The new obstetrics dataset is representative of the broader domain and that manual correction of hallucinations provides a fair, scalable basis for comparing model performance.","pith_extraction_headline":"Supervised clinical section segmentation models drop in performance when moving from MIMIC-III to obstetrics notes, while zero-shot models remain robust after correcting for hallucinated headers."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2602.17513/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":24,"sample":[{"doi":"","year":2021,"title":"Bridging the Domain Divide: Supervised vs. Zero-Shot Clinical Section Segmentation from MIMIC-III to Obstetrics","work_id":"ee048cf4-d010-48bb-82ab-878826013785","ref_index":1,"cited_arxiv_id":"2602.17513","is_internal_anchor":true},{"doi":"","year":null,"title":"ONC serves as a realistic benchmark for studying section segmentation in underexplored clinical subdo- mains and is intended for community reuse","work_id":"b73dc831-2aa8-45d8-b8c0-c0868edd642a","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Domain-Specific Evaluation of Supervised Models:We assess whether transformer- based supervised models originally trained on public datasets can effectively generalize to obstetrics notes. By comparin","work_id":"47784246-752b-4e2f-b0be-ff9762b50c18","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Our exper- iments reveal challenges (e.g., hallucinated section headers) as well as the potential ben- efits of zero-shot strategies, especially when annotated data are scarce","work_id":"5eba15b0-04d0-4ec6-8f68-5343728be72a","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2008,"title":"Related Work Before the emergence of advanced machine learn- ing and NLP techniques, early approaches to clin- ical section segmentation primarily relied on rule- based methods. Denny et al. (2008), f","work_id":"bcadcf55-8118-41dd-847c-788503519ce6","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":24,"snapshot_sha256":"f758c4cf6c8f2a5907b749623bce36c2fd023bd4213a9ff7564c9be6cd31ca33","internal_anchors":4},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}