P CLINICAL AI

 A

Domain	Key Insight	Evidence from Document
Benchmark saturation	LLMs now exceed 90% on medical exams, making traditional tests less useful.	“Newer top models are routinely exceeding 90% accuracy… we are quickly losing the ability to evaluate them effectively.”
Clinical reasoning gap	High exam scores ≠ real‑world reasoning; models misread context, anchor early, and fail to ask follow‑ups.	“Performing well on a test is not always the same as performing well in the real world… failing to ask the right follow-up questions.”
PrIME‑LLM metric	New scoring system evaluates 5 domains: differential, testing, final diagnosis, management, misc reasoning.	“PrIME‑LLM calculates a normalized polygonal area across five domains…”
Overall accuracy vs balanced reasoning	Raw accuracy 81–90%, but balanced PrIME‑LLM scores only 64–78%.	“Models ranged from 64% to 78%, showing more separation than the narrow 81–90% accuracy band.”
Differential diagnosis failure	Catastrophic weakness: >80% failure rate across all models.	“Failure rates exceeded 80% across all 21 models tested. Not some of them, all of them.”
Premature closure	Models collapse onto one answer instead of holding multiple possibilities.	“LLMs appear to collapse prematurely onto a single answer rather than preserving uncertainty…”
Analogy: pizza vs alternatives	LLMs identify the obvious diagnosis but fail to generate alternatives.	“They detect tomato… and they say ‘pizza.’… but a skilled chef would also consider what else it could be.”
High‑stakes risk	Missing alternatives in medicine = missed cancer, delayed intervention.	“Anchoring… is not a lost round—it is a missed cancer, a delayed intervention…”
Baseline, not ceiling	Study used off‑the‑shelf models without tools, search, or structured reasoning.	“No real-time search… no structured reasoning workflows… this was a baseline evaluation.”
Reasoning‑optimised models	These models scored higher (76% vs 67%), showing architecture matters.	“Reasoning models scored significantly higher… mean: 76%… nonreasoning models: 67%.”
Human vs AI performance	LLMs can outperform clinicians on structured cases, but not proven to improve clinician decisions.	“GPT‑4 outscored both attending physicians… but giving physicians access to an LLM… did not meaningfully improve their performance.”
Clinical deployment caution	Safe for low‑stakes tasks; unsafe for autonomous diagnostic reasoning.	“For lower-stakes… the case for adoption is reasonable… Autonomous diagnostic reasoning is a different proposition entirely.”
Core risk: false confidence	Models project confidence where uncertainty is required.	“These models may project confidence precisely where clinical reasoning demands uncertainty.”
Bottom line	Until prospective real‑world evidence exists, caution is essential.	“Until we have prospective data… we should continue exercising caution…”