Clinical AI's new pitch isn't accuracy — it's proof it matches a human

Matching the human

The clearest version of the new pitch comes from Komodo Health, which says it ran a blinded study comparing its Marmot AI platform against a human analyst for real-world evidence generation. The study focused on a non-ambulatory Duchenne muscular dystrophy cohort, and both pipelines produced identical counts at all eight cohort selection steps. That is the kind of result a buyer can check: not a score on a hidden test set, but a step-by-step match with the person the tool is supposed to assist. Komodo Health also reports that the initial cohort size was 30,221 patients and the final cohort size was 1,130 patients, so the match held through substantial narrowing. m:…6476

The study is more persuasive because it reports where the match broke. Age, region, and race and ethnicity distributions matched exactly between the two pipelines, and 16 of 17 Charlson Comorbidity Index components matched, but the mean comorbidity score was 1.5 in the manual analysis and 1.1 in the Marmot analysis, and payer mix differed between the two pipelines. Komodo Health says the discrepancies were traced to derived-variable implementation, which is another way of saying the two sides defined a computed variable differently. Three repeated Marmot executions returned identical outputs under the tested conditions, which speaks to repeatability, but the company also says validation of downstream analyses, including healthcare resource utilization and cost outcomes, is ongoing. The honest boundary is part of the pitch: the tool matched a human on cohort selection, and the rest is not yet proven. m:…6476

Explaining itself, by source

A preprint on in-hospital detection of lung cancer and heart failure makes a parallel argument for a different kind of proof. Multimodal fusion of routine electronic health record data, combining lab results, radiology reports, and ECGs, significantly improves in-hospital detection of lung cancer and heart failure compared with laboratory data alone. The study also reports that FedProx federated training recovers 82 percent of centralized performance gains for rare outcomes like lung cancer, outperforming FedAvg, which matters for hospitals that cannot pool patient data centrally. Detection improves when sources are combined, and training can stay distributed without giving up most of the gain. m:…5821

But the preprint adds a check that many accuracy reports skip: explanation faithfulness varies by data source, with laboratory-based explanations failing weight-randomization checks. An explanation that does not survive a randomization check is not really explaining the model's decision, even if it looks plausible. The implication for buyers is that explainability is not one property of a model but a separate test for each data source feeding it, and a model can be explainable for radiology while misleading for labs. m:…5821

A second preprint, on classifying Rett syndrome from resting-state EEG, shows the same discipline applied to model selection. Random Forest achieved the best performance among the classical machine learning models, EEGNet achieved the highest deep learning performance, and 4-second EEG epochs offered the optimal trade-off between model performance and robustness. Delta and theta oscillations, along with centro-frontal brain regions, are key contributors to classification, which gives clinicians a physiological story to check rather than a black-box verdict. The pattern across both preprints is consistent: the result worth reporting is not only which model won, but under what conditions and for which signals. m:…5407

Proof as a product

Tempus shows the commercial side of the same shift. The company says it entered a multi-year agreement with Bristol Myers Squibb under which Bristol Myers Squibb will license multimodal oncology data from Tempus through 2031 and gain access to the Lens Platform. Tempus says the licensed data supports biological and clinical insight identification, novel drug target discovery, and therapeutic development across the oncology portfolio, and that the Lens Platform allows researchers to train and execute advanced AI models. The company also says the existing relationship includes applying AI and real-world data to optimize clinical trial design and neuroscience initiatives, so the new agreement expands a working relationship rather than starting one. m:…5620

Put the three items together and the market's new question comes into focus. Buyers are no longer asking only how accurate a clinical AI tool is; they are asking whether it matches a human on the same task, whether its explanations hold up per data source, and whether its data and platform are licensed as durable infrastructure. Tempus says its agreement expands a long-standing strategic relationship, and Komodo Health is careful to say downstream validation is ongoing. That combination, confidence about what was proven plus clarity about what was not, is becoming the price of admission. Vendors that can only quote a score will find the conversation has moved on to concordance, and concordance is a harder test to pass. m:…6476 m:…5620