Risk scores built on diagnosis codes alone stopped predicting death and flipped hospital rankings

A problem-list comorbidity score showed no link to death in 1,596 veterans. In Finland, a condition that doubles mortality looked protective in codes…

Oct 31 • David Talby

22

3

1

Ten questions to ask before you trust a clinical AI benchmark

Sample sizes, paired tests, who's asking, subgroups, clinician review, and private case text: what a safety or bias benchmark needs before its score…

Oct 26 • David Talby

22

4

3

Building AI software got cheap and proving it works became the job

The development team has become the QA team. It has five things to prove.

Oct 10 • David Talby

28

6

1

September 2026

Ten regulatory duties for clinical AI that no benchmark tests today

The 619-instrument Q3 2026 Policy Suite and the ten properties of a benchmark grounded in it

Sep 19 • David Talby

49

7

2

Healthcare analytics built on structured data alone miss almost 40% of documented diagnoses

HEDIS rates from claims alone ran 20 points below chart-reviewed rates, Medicare claims found 2% smokers where the survey found 10%, and only 3% of…

Sep 12 • David Talby

24

3

10

Fact-level provenance in healthcare AI: the 42 capabilities behind an FDA-ready clinical data platform

The FDA’s December 2025 real-world evidence guidance requires provenance and accuracy at a per-fact level. Here is the full capability inventory, the…

Sep 5 • David Talby

19

2

2

There is no partial credit in de-identification

New privacy tools score between 0.55 and 0.91 PHI F1 on clinical notes; a purpose-built pipeline holds 0.98 recall, above the two-expert consensus line…

Sep 1 • David Talby

24

2

August 2026

Benchmarking AI impersonation of licensed professionals: 11 laws and 380 adversarial tests

A new red-team benchmark measures whether frontier models impersonate doctors, lawyers, and financial advisors, and whether they disclose being AI…

Aug 8 • David Talby

58

1

The cognitive bias problem LLMs inherited from doctors

Take a clinical note about a 52-year-old with chest pain, and ask a frontier model what to do.

Aug 1 • David Talby

84

3

4

July 2026

Small, Private, and First on All Fifteen: The New Medical LLM Benchmark Results

First on all 15 clinical and biomedical benchmarks, averaging 80.9 against the newest frontier releases, running on a single GPU inside your own…

Jul 25 • David Talby

52

2

What we learned building Medical LLMs that academic medical centers trust

Originally published alongside the 2025 InfoWorld Technology of the Year Award announcement, December 2025.

Jul 15 • David Talby

1

When the Title Outruns the Study: General-Purpose vs. Healthcare-Specific AI

A Brief Communication in Nature Medicine made the rounds last month under a headline that is hard to misread: “General-purpose large language models…

Jul 12 • David Talby

37