SIFTING: traceable LLM extraction for lung-cancer T-staging

Self-hosted Llama extracts structured T-stage from radiology reports with source-text links — 90% accuracy vs a four-expert reference on 130 reports.

Primary source: arXiv:2609.07185: https://arxiv.org/abs/2609.07185

What’s new: Clinical NLP demos often hand an entire radiology report to a big model and hope the answer comes back clean. SIFTING (arXiv:2609.07185, submitted 7 Sep 2026) is a framework aimed at the failure mode that blocks adoption: unstructured, untraceable outputs. It combines segment-level processing with structured prompts and strict output control so each extracted finding is linked back to source text. The authors apply it to tumor T-staging from lung-cancer radiology reports (SIFTING-T-stage), running a 4-bit quantized open-source Llama-3.3-70B (~35 GB) in a fully self-hosted setup.

On 130 reports, scored against a reference standard built by four clinical experts and compared with conventional single-prompt LLM baselines (bootstrap confidence intervals), SIFTING-T-stage reaches 90% accuracy (95% CI: 84–95). The paper reports performance comparable to much larger state-of-the-art LLMs with reasoning capabilities, statistical interchangeability with the expert panel on this task (p < 0.001), and full traceability via source-text references.

Why it matters: Staging from free text is exactly the kind of “AI beyond chat” work clinics actually need — structured fields that can enter a workflow, with an audit trail a radiologist or oncologist can check. Self-hosting matters here for the usual reasons: patient text stays under local control, the model version is pinned, and the extraction recipe is reproducible. Linking answers to spans is not a nicety; it is what makes validation possible instead of trusting a black-box label.

For Bitware’s beat, this is a methods-plus-medical signal: not a new foundation model, but a concrete extraction stack (segmentation, constrained outputs, provenance) evaluated on a real clinical document type.

Caveats: This is a preprint, not a peer-reviewed trial. The demo is a single task (T-stage) on 130 reports — useful as a controlled study, not a multi-institution claim. “Interchangeable with clinical experts” is a statistical result on this labeled set; it is not permission to run unsupervised staging in care. A quantized 70B still needs serious hardware and ops discipline. And accuracy on T-stage does not automatically transfer to N/M staging, other cancers, or messier multi-site report styles. Treat it as evidence that traceable, self-hosted extraction can match strong baselines on a narrow, expert-scored task — then ask for external validation before anyone calls it production-ready.