Instruction-Tuning Large Language Models with Medical Classification Data Improves ICD Coding of German Tumor Diagnoses

Instruction-Tuning Large Language Models with Medical Classification Data Improves ICD Coding of German Tumor Diagnoses

The authors report that instruction-based fine-tuning on public medical classification data improved open-weight LLMs’ coding of German tumor diagnoses. Their motivation is practical: smaller models may support privacy-preserving automation, but they often struggle with coding accuracy in German-language contexts. On coded diagnoses from a local tumor documentation system, ICD-10-GM accuracy increased from 1.4–24% to 41–58%; partial accuracy, counting three-character codes, increased from 31–74% to 73–83%. ICD-O-3 topography coding also improved, but remained considerably less accurate. Malformed code outputs fell to 0% for all models, and tumor-diagnosis recognition reached 99%. The authors also report that accuracy correlated positively with model size, while fine-tuning narrowed the gaps between small and large models. In their comparison, Qwen3’s reasoning mode generally performed worse than fine-tuning and was over 100 times slower. The results suggest that public catalogues can provide training material for this documentation task, but the reported evaluation is tied to the local test data and does not establish performance in other settings.

Publication: 6d248e78-4b39-46e4-89e9-1b7698cbbc48

attribution — pass: The reading attributes the results and motivation to the paper's authors with phrases such as 'The authors report' and 'In their comparison.' It does not present the source's claims as established fact or as Escape Meridian's view.

Authors:Stefan Lenz, Lakisha Ortiz Rosario, Georg Vollmar, Arsenij Ustjanzew, Fatma Alickovic, Thomas Kindler, Torsten Panholzer

summary — uncertain: The figures are faithful, but a lay reader is not told what ICD coding is (assigning standard diagnosis codes, not programming). ICD-10-GM, ICD-O-3 topography, open-weight and instruction-tuning also need tooltips. The summary drops the authors' estimated ceiling of 60-79% exact and 81-94% partial, which changes how the 41-58% result reads. It also omits the 500,000+ question-answer pairs built from public catalogues and the released dataset and checkpoints. Smallest repair: add one plain sentence defining ICD coding and one stating the estimated ceiling.

the upper limit for ICD-10 coding performance was estimated at 60-79% for exact and 81-94% for partial (three-character codes only) derivation

date — pass: originalPublished is 2025-10-15, which matches the v1 submission date. The 5 Nov 2026 date is a later revision and the feed date is later still, so neither should be displayed.

Submitted on 15 Oct 2025 (v1), last revised 5 Nov 2026 (this version, v2)

support — pass: Each claim matches its abstract excerpt. The comparison of ICD-O-3 with ICD-10-GM follows from 'started and remained considerably lower.' The limitation about local test data and the question about generalization are framed as unknowns rather than as negative claims. The closing inference about public catalogues mirrors the authors' own stated potential.

Our findings highlight the potential of leveraging public catalogues to build instruction datasets that improve LLMs in medical documentation tasks.

media — pass: The selected media is the verified arXiv abstract page for this paper. It is the direct original source, it is relevant, and it is the only supplied candidate.