Guy's and St Thomas' Refined a Rules-Based Decision Support Platform Until It Matched Its Own Breast MDT on 250 Cases

Five cohorts of 50, entered blind

Patients discussed over the preceding 16 months were entered in blocks of 50, blinded to the treatment recommendation the team had made, and after each block the platform's output was compared against the recorded MDT decision. The first benign block produced nine discordant cases out of 50. All nine traced to benign entities that had never been encoded, and the authors list them: fat necrosis, scarring, sclerosing adenosis, dermatitis and eczema, intraductal papilloma, inflammation, lactational change, lipoma, hamartoma and fibromatosis. Those were added, the second benign block ran clean, and no third benign cohort followed.

The cancer arm did not move in a straight line. The first malignant block, restricted to T1-2 N0-N1 M0 disease, came in at 94% with three discordances. The governance group then widened the criteria to all primary breast cancer short of recurrent or metastatic disease and loaded in comorbidities, ECOG performance status, treatment contraindications, reconstruction, radiotherapy fields, neoadjuvant chemotherapy and open trials at the center. The next block came back at 92%, four discordances, below the first, and only the third cohort reached 100%.

Knowledge representation with a class IA approval

The authors place Deontics in the knowledge representation branch of artificial intelligence, meaning encoded guideline logic that produces deterministic output and can show the rule behind each recommendation. They note the platform carries class IA device approval from the Medicines and Healthcare products Regulatory Agency, and they contrast it with Watson for Oncology, which used machine learning, and NAVIFY, which used rule-based process management. The encoded guideline set began with a benign breast disease protocol, NICE NG101, and Royal College of Radiologists and Pathologists guidance, and grew as validation exposed gaps.

No stopwatch was involved

The stated aim is to free MDT time for complex cases, and no time was measured. Using the original conservative criteria, 31% of benign and 44% of malignant patients would still have required formal discussion. After the update, the same calculation gives 31% of benign patients and 1% of malignant ones.

That second figure is a projection from retrospective agreement, not an observed reduction in caseload, and it excludes the recurrent and metastatic patients the design carved out from the start. On what the work adds, the authors write that artificial intelligence-based clinical decision support systems "can accurately triage cases directly without requiring full multidisciplinary team discussion, thereby prioritising time for discussion of more complex cases." Their own limitations section is narrower. The standard of care was validated internally, against past decisions of the same team, and they acknowledge that the absence of external validation limits generalizability, with prospective concordance tracking still to come.

Patkar was excluded from design, selection and analysis

The competing interests statement handles Patkar's dual role at length. It records that "to safeguard analytical independence, the role of VP was strictly limited to providing technical support: implementing the Standard of Care (SoC) into the Clinical Decision Support (CDS) platform, updating the CDS platform to incorporate SoC refinements, and providing extracted data from the platform." He was excluded from study design, case selection, cohort division, data entry and analysis. Hannah Jeffery and Aaditya Prakash Sinha are joint first authors, and Arnie Purushotham of King's College London is corresponding author.

Why This Matters to the APO|APE Reader

Hammer and colleagues, cited by this team, measured a 30% cut in MDM preparation time with NAVIFY across four cancer categories, and Khattak and colleagues reported a 34% reduction in prostate MDM caseload in 2025. Patkar's own earlier breast work reported 97% concordance with a 61% increase in trial-eligible patients identified. No equivalent figure exists for this build, and it is the first thing a procurement panel will ask for. What transfers is the shape of the effort: a cohort-by-cohort audit in which the tool is edited to match the team, with the pathology vocabulary it was missing supplied by the cases it got wrong.