PaperlayerView exact revision on arXiv

Source-linked AI summary

Development and Feasibility Evaluation of an Edge AI as Medical Device System for Breast Cancer Multidisciplinary Team Meetings

Aarzoo Dhiman, Farzana Haque, Kartikae Grover, Lydia Brian Smith, William Stephen Jones

arXiv:2608.22108v1cs.AIcs.LG

TL;DR

Breast cancer MDTs need to integrate complex clinical information under time pressure, while existing systems often depend on cloud processing or curated case text. This study develops and bench-tests a fully on-device pipeline combining ASR, locally hosted LLMs, and guideline-informed retrieval for MDT documentation and recommendations. The system approached a commercial clinical ASR benchmark on augmented audio and MedGemma-RAG identified more MDT-concordant interventions than a proprietary cloud comparator, but prospective clinical evaluation remains outstanding.

  • Problem

    Breast cancer MDTs face time and information-integration pressures, while prior decision-support studies largely use curated text and cloud-hosted models rather than live discussions.

  • Method

    The study develops and evaluates an edge-deployed pipeline integrating ASR, locally hosted LLMs, and retrieval-augmented generation for structured MDT information and NICE-informed recommendations.

  • Results

    MedGemma-RAG identified 2.3 times more MDT-concordant interventions than a proprietary cloud comparator, while ASR performance was within 1.58% WIL of a commercial clinical benchmark on augmented audio.

  • Takeaways & Limitations

    The findings support the feasibility of privacy-preserving, fully on-device AI for MDT documentation and guideline-informed decision support.

  • Takeaways & Limitations

    The study provides analytical validation and bench feasibility testing rather than clinical evaluation, based on six analysable cases and two authentic MDT recordings.

Abstract

from arXiv · show

Breast Cancer Multidisciplinary Team (MDT) meetings manage increasingly complex cases under considerable time pressure, and documentation requirements can reduce clinical efficiency and decision quality. Existing AI based MDT workflows rely on cloud-based processing, limiting their use because patient discussions contain identifiable information. We developed a fully on-device AI pipeline using open-source Automatic Speech Recognition (ASR) and Large Language Models (LLMs) that transcribes breast cancer MDT discussions, structures clinical information, and generates treatment recommendations using retrieval-augmented generation (RAG) grounded in National Institute for Health and Care Excellence (NICE) guidance. The pipeline runs on a single NVIDIA Jetson AGX Orin, ensuring that patient audio, transcripts, and outputs remain within institutional infrastructure. Evaluation included two recorded simulated MDT discussions, ten clinically validated synthetic discussions, and 1,270 acoustically augmented recordings. Optimisation of Whisper large-v3 reduced word error rate by 20.7% and 24.4% on the recorded discussions and achieved performance within 0.58% WER and 1.58% word information lost of a commercial clinical ASR benchmark on augmented audio. MedGemma-RAG identified 2.3 times more MDT-concordant interventions than a proprietary cloud comparator (p = 0.020), with no significant difference in overall accuracy. Stakeholders identified automated documentation, treatment recommendation support, and case triage as the most credible near-term applications while highlighting workflow integration, governance, and clinician trust as key implementation challenges. These findings demonstrate the feasibility of privacy-preserving, fully on-device AI for MDT documentation and guideline-informed decision support, providing a foundation for prospective clinical evaluation.

1. Introduction

Breast cancer MDTs must integrate complex, distributed clinical information under time pressure, while existing AI workflows often rely on cloud processing and curated case text. This study develops and evaluates an open-source, edge-deployed AIaMD pipeline that processes MDT discussions locally for transcription, structured extraction, and NICE-informed recommendations.

  • Clinical motivation: MDT meetings review growing caseloads in limited time, creating decision fatigue and reducing information quality and team contribution.The relevant evidence describes only a few minutes per patient and measurable cognitive and operational challenges.
  • Clinical motivation: Breast oncology recommendations require detailed clinicopathological information and guidance distributed across multiple NICE guidelines and technology appraisals.This creates a gap between decision-relevant information and what can realistically be synthesised during meetings.
  • Research gap: Existing MDT decision-support studies predominantly use structured referral letters or curated case summaries rather than live, conversational MDT discussions.Consequently, they do not address speech-derived clinical information or end-to-end workflows from discussion to recommendation.
  • Research gap: Most reported systems rely on proprietary cloud-hosted models, raising governance, confidentiality, cost, reproducibility, and information-loss concerns for healthcare deployment.External inference and intermediate structured summaries can introduce additional opportunities for information loss.
  • Study contribution: The proposed AIaMD pipeline combines open-source ASR and LLMs to capture MDT discussions, structure clinical information, and generate NICE guideline-informed recommendations.The pipeline runs locally on a single edge device using quantised models, without external data transmission.
  • Study contribution: The study evaluates analytical validation and bench feasibility, while prospective clinical evaluation is reserved for future studies.The assessment includes authentic simulated, synthetic, and acoustically augmented audio, with local processing supporting data locality and NHS information governance.

3. Results

The results evaluate ASR robustness and optimisation, model selection, treatment prediction, and stakeholder priorities for AI-assisted breast cancer MDT workflows. Optimised Whisper large-v3 improved transcription, while MedGemma-RAG produced more clinically appropriate interventions under some evaluation conditions without significantly different overall accuracy.

  • ASR robustness under acoustic degradation: Across 1,270 augmented recordings, WER remained low below WIL 0.4 but variability increased sharply at higher WIL values.WIL values of 0.4–0.6 corresponded mainly to reverberation and partial speaker overlap; values above 0.6 produced wider interquartile ranges and more extreme outliers.
  • ASR robustness under acoustic degradation: ASR failures at WIL values above 0.6 were intermittent and associated mainly with overlapping speech, speed perturbation, and reverberation.The findings indicate that interacting temporal and multi-speaker distortions affected performance more than individual acoustic perturbations alone.
  • Model selection: Whisper large-v3 recovered more curated breast cancer diagnostic and treatment phrases than Parakeet, with recall 0.56 versus 0.52.Parakeet had the lowest WER, whereas Whisper large-v3 had the lowest WIL, motivating its selection for subsequent optimisation.
  • ASR optimisation: WER fell to 0.3635 and WIL to 0.4237 with DNSMOS threshold 3.5, narrowing the gap with the commercial benchmark to 0.58% WER and 1.58% WIL.These represented improvements of 7.3% and 4.5%, respectively, over the default configuration.
  • ASR optimisation: On held-out authentic recordings, tuned decoding reduced WER to 0.49 and 0.2070, corresponding to reductions of 20.7% and 24.4%.The lowest WER for Recording 2 was 0.19, a 31.4% reduction, with threshold 3.0; thresholds of 3.0–3.5 performed consistently best.
  • Model selection: Clinical terminology recall did not consistently improve with aggregate transcription accuracy: silence removal improved WER but reduced recall from 0.31 to 0.22.The findings show that aggregate transcription metrics and preservation of clinically important terms did not necessarily improve in parallel.
  • LLM-based treatment prediction: MedGemma produced more complete and consistently structured case summaries than Palmyra-Med, whose outputs often omitted information from longer discussions.MedGemma was selected for subsequent recommendation generation under edge deployment constraints.
  • LLM-based treatment prediction: Under FT2, MedGemma-RAG achieved higher precision (0.44 vs 0.42), Jaccard coefficient (0.36 vs 0.34), and clinically appropriate additional interventions (0.53 vs 0.47) than ChatGPT-5.2.It also reduced overgeneration from 0.58 to 0.56 and hallucination from 0.12 to 0.06.

4. Discussion

The discussion supports the technical feasibility of a privacy-preserving, edge-deployed AI pipeline for breast cancer MDT documentation and decision support, while identifying clinical, workflow, and evaluation boundaries. Its performance depends on preserving clinically important terminology, grounding recommendations in NICE guidance, and validating the system prospectively in live practice.

  • Technical feasibility: The edge pipeline combines spontaneous MDT speech processing with structured summarisation and guideline-informed treatment recommendation while keeping identifiable patient data within institutional infrastructure.Unlike prior tumour-board studies using curated case summaries, it processes spontaneous MDT discussions.
  • Clinical limitations: Transcript-derived summaries cannot capture the full clinical context available to MDT members, and ASR errors may propagate into downstream tasks.The LLM nevertheless frequently generated clinically coherent case summaries, suggesting partial compensation for imperfect transcripts; future work should detect clinically significant errors beyond aggregate ASR metrics.
  • ASR performance: 0.3727 versus 0.3927 WER gave NVIDIA Parakeet a marginal lexical advantage over Whisper large-v3, whereas Whisper recovered more clinically relevant terminology.In breast cancer MDTs, receptor status, laterality, and disease stage can matter more than overall lexical accuracy; some WER-improving configurations reduced terminology recall.
  • Recommendation generation: Retrieval of relevant NICE recommendations improved agreement with MDT decisions and reduced inappropriate recommendations compared with generation without RAG.MedGemma-RAG showed higher precision, Jaccard agreement, positive-class performance, and lower hallucination rates, while ChatGPT-5.2 had higher recall and more additional interventions.
  • Implementation: Interoperability, governance, and workflow integration remain necessary for deployment despite reduced data-transmission risks from on-device processing.Stakeholders identified documentation support and case triage as credible near-term applications, while implementation requires integration across electronic records, radiology, pathology, and MDT systems.
  • Study limitations: The study provides analytical validation and bench feasibility testing rather than clinical evaluation, with small authentic datasets, clustered observations, equal intervention weighting, and unassessed calibration.Synthetic speech did not reproduce the disfluencies and spontaneous turn-taking of authentic MDT discussions, and prospective live-MDT evaluation is planned.

5. Conclusions

The study demonstrates feasibility of an edge-deployed AI pipeline for breast cancer MDT documentation and guideline-informed treatment support without transmitting identifiable patient data beyond institutional infrastructure. Its performance and stakeholder findings support further clinical evaluation, while authentic conversational variability and implementation requirements constrain routine use.

  • System contribution: The open-source edge pipeline integrates ASR, locally hosted LLMs, and guideline-informed retrieval for breast cancer MDT documentation and treatment recommendation.The system keeps identifiable patient data within institutional infrastructure.
  • Technical findings: 1.58% WIL separated the optimised ASR pipeline from a commercial clinical benchmark on augmented audio.The benchmark was used for comparison rather than as part of the proposed pipeline.
  • Decision support: 2.3 times more MDT-concordant interventions were identified by MedGemma-RAG than by a proprietary cloud-based comparator.The comparison also reported consistently low hallucination rates.
  • Evaluation boundary: Performance was more variable on authentic MDT conversations than on synthetic audio, underscoring the importance of evaluating conversational clinical data.
  • Implementation: Documentation support and case preparation were identified as the most credible near-term applications, while workflow integration and governance remain implementation challenges.
  • Next steps: Prospective clinical evaluation, consequence-weighted assessment, and post-deployment monitoring are required before routine clinical use.

Supplementary Information

The supplementary information documents the literature-search strategy and prior breast cancer MDT AI systems, including their roles, study designs, and reported outcomes.

  • Search strategy: The literature search used PubMed/MEDLINE and combined breast cancer MDT terms with artificial intelligence, clinical decision support, LLM, NLP, and RAG terms.
  • Search strategy: A parallel search addressed automatic speech recognition studies in breast cancer MDT settings.
  • Related systems: Supplementary Table 1 summarises published AI systems for breast cancer MDT decision support, including system role, study design, evaluation approach, and outcomes.
  • Related systems: The reviewed evidence includes a meta-analysis of 9 studies involving 2,463 patients and retrospective evaluations of complex breast cancer cases.

Supplementary Section 2: Datasets, Models and Full Configuration Results

The supplementary methods describe synthetic MDT data generation, acoustic augmentation, evaluated ASR systems, and Whisper large-v3 optimisation for adaptive transcription.

  • Synthetic datasets: Ten synthetic MDT discussions represented varied durations and multidisciplinary speaker-role combinations.Roles included radiologists, pathologists, surgeons, oncologists, breast care nurses, coordinators, chairs, and palliative care specialists.
  • Synthetic datasets: 1,270 augmented audio files were produced by applying 127 variants per recording across acoustic stress-testing configurations.The augmentations included reverberation, overlapping speech, resampling, pitch shifting, environmental noise, and dropout.
  • ASR systems: The ASR evaluation included open-source Whisper and GPU-optimised models, WhisperX alignment and diarization, and a commercial clinical benchmark.
  • ASR systems: The commercial clinical ASR service required cloud processing and ongoing operational cost, so it was used as a benchmark only.
  • Optimisation: Whisper large-v3 was optimised with Optuna over 100 trials to minimise WER on Tier 3 augmented audio.The search varied decoding parameters including temperature, beam size, length penalty, and hallucination-silence threshold.
  • Optimisation: Supplementary Table 7 reports WER and WIL for default and optimised Whisper large-v3 configurations on two held-out authentic MDT recordings.

Supplementary Section 3: ASR Evaluations

The supplementary ASR evaluation defines transcription metrics, assesses curated clinical phrase recall, and documents the synthetic case content and matching limitations.

  • Evaluation metrics: WER is defined from substitutions, deletions, and insertions divided by the number of words in the reference transcript.
  • Clinical phrase evaluation: Curated breast cancer diagnostic and treatment phrases were matched against ASR output using exact and fuzzy 1–6-gram comparison.Fuzzy matching used Levenshtein similarity of at least 80%.
  • Clinical phrase evaluation: Recall was computed as TP / (TP + FN), while precision was not computed because candidate terms were restricted to a predefined vocabulary.
  • Clinical scenarios: The evaluated cases covered screening lesions, male breast cancer, multifocal and triple-negative disease, metastatic disease, DCIS, and postoperative treatment scenarios.
  • Clinical scenarios: The phrase lists included clinical findings, pathology, biomarkers, treatment options, surgery, supportive care, and referral decisions.
  • Evaluation caveats: Corrected spellings and fuzzy matching thresholds were relevant because matching was sensitive to spelling variants.

1. Consider the given text as the MDT meeting discussion transcript of a single

The workflow extracts MDT case information and treatment decisions from transcripts, preserves explicit reasoning and uncertainty, and aligns proposed strategies with NICE guidance. It uses retrieval-augmented generation to supply relevant guideline context for treatment planning.

  • The system corrects ASR-derived medical terminology while preserving the transcript’s information and structure.
  • Case extraction covers demographics, diagnosis, pathology, imaging, nodal status, and biomarkers without adding information absent from the transcript.
  • Treatment extraction reconstructs MDT decisions, including sequencing, conditional plans, referrals, and explicit clinical reasoning.
  • The system is intended for clinician-reviewed documentation and decision support rather than autonomous medical advice or recommendation.
  • The proposed strategy is checked against relevant NICE guidance, using retrieved guideline context from a FAISS vector database.
  • The workflow uses conditional language and explicit IF/THEN branches for pending biomarker results, while retaining decisions that reject standard steps.

1. Understand the case description carefully and the context from the NICE

The labelling procedure uses only information present in the transcript and NICE context to classify treatment options as YES, NO, or NEI. It preserves explicit MDT decisions and avoids unsupported pathway assumptions.

  • Treatment options are classified as YES, NO, or NEI, with a brief reason for each label.
  • The full predefined treatment-option list is labelled without adding or removing options.
  • YES requires an explicitly planned, recommended, accepted, or requested intervention aligned with NICE guidance or established knowledge.
  • NO applies when an intervention is explicitly rejected, ruled out, or contrary to the discussed management plan.
  • NEI is used when the case description or NICE guidance lacks enough information to choose an option.
  • The procedure preserves explicit MDT rejections, uses conditional wording, and represents pending biomarkers with IF/THEN treatment branches.

Supplementary Section 5: LLM Decoding Hyperparameters

The supplementary section specifies model inference settings, intervention taxonomy, evaluation questions, metrics, and paired comparisons between MedGemma-RAG and ChatGPT-5.2. Results show significant advantages for selected class-specific metrics but no significant overall-accuracy differences.

  • LLM Decoding Hyperparameters: Inference parameters were specified separately for case extraction, treatment separation, and guideline-grounded treatment generation.
  • LLM Decoding Hyperparameters: Guideline-grounded treatment generation used temperature 0.1, top_p 0.9, repeat_penalty 1.1, top_k 40, and max_tokens 1000.
  • Evaluation: The model classified a full intervention taxonomy and was evaluated for MDT-plan matching, clinical validity, completeness, and hallucination.
  • Results: 83.6% versus 80.8% overall accuracy was observed for Scenario A, without a significant difference between MedGemma-RAG and ChatGPT-5.2.
  • Results: 60.1% versus 61.4% overall accuracy was observed for Scenario B, also without a significant difference.
  • Results: Scenario A YES precision, recall, and F1 were significantly higher for MedGemma-RAG than ChatGPT-5.2.

Supplementary Section 8: DECIDE-AI Alignment and Deferred Items

The study is positioned as analytical validation and bench feasibility testing aligned with DECIDE-AI, while live clinical evaluation remains deferred. Unassessed calibration, workflow integration, clinical safety, users, and patient outcomes define the main implementation boundaries.

  • Evaluation status: The study reports analytical validation and bench feasibility testing rather than prospective clinical evaluation.
  • Intended use: The intended use is documentation support and case triage under clinician review, not autonomous recommendation.
  • Deferred items: Calibration was not assessed and is identified as required before recommendation-facing use.
  • Deferred items: No users operated the system, and live clinical integration was deferred pending live evaluation.
  • Safety: Hallucination and overgeneration were measured on bench data, while clinical-use safety requires prospective study.
  • Deferred items: Patient-related outcomes were unavailable because no patients were involved and no clinical outcomes could be measured.
Loading 2608.22108v1…