When Artificial Intelligence Plays It Safe: Large Language Models Preserve Surgical Sensitivity at the Cost of Diagnostic Specificity in Acute Abdominal Pain

BackgroundIn acute abdominal pain, artificial intelligence may appear safe by preserving surgical sensitivity while increasing resource use through reduced specificity. We evaluated this safety-stewardship trade-off across frontier large language models (LLMs) to provide an exploratory diagnostic benchmark.MethodsWe performed a cross-sectional in-silico study using 50 standardized synthetic vignettes (25 urgent surgical, 25 non-urgent) internally developed by our investigative group. GPT-5.5, Claude 4.0 Opus, and Gemini 3.0 Ultra were evaluated using identical zero-shot deterministic prompting against an internally established, blinded dual-specialty expert consensus.ResultsExact primary-diagnosis accuracy was 92.0% (95% CI, 81.2-96.8%) for GPT-5.5, 88.0% (95% CI, 76.2-94.4%) for Claude 4.0 Opus, and 76.0% (95% CI, 62.6-85.7%) for Gemini 3.0 Ultra. Surgical sensitivity remained high across models (84.0% to 96.0%), whereas specificity was lower (64.0% to 80.0%) than specialist consensus. Models demonstrated a conservative escalation pattern with frequent over-triage decisions. Positive likelihood ratios ranged from 2.33 to 4.60. Over-investigation bias was the predominant error pattern.ConclusionIn this exploratory benchmark, frontier LLMs consistently prioritized surgical sensitivity over specificity in internally authored structured acute-abdomen vignettes. These findings reflect differences in performance on controlled synthetic scenarios rather than independently validated clinical diagnostic accuracy. Claims regarding clinical patient safety or readiness for implementation require future prospective validation with independent, real-world clinical cases and external multidisciplinary adjudication.

Authors

Institutions

Publication Details

Journal
The American Surgeon
Published
2026-11-20
DOI
https://doi.org/10.1177/00031348261508223
Primary Topic
Artificial Intelligence in Healthcare and Education
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

When Artificial Intelligence Plays It Safe: Large Language Models Preserve Surgical Sensitivity at the Cost of Diagnostic Specificity in Acute Abdominal Pain

Fatih Başak, Sıbel Ocak Serın
4 links
The American Surgeon
Artificial Intelligence in Healthcare and Education
article

When Artificial Intelligence Plays It Safe: Large Language Models Preserve Surgical Sensitivity at the Cost of Diagnostic Specificity in Acute Abdominal Pain

Fatih Başak, Sıbel Ocak Serın
article en

Abstract

BackgroundIn acute abdominal pain, artificial intelligence may appear safe by preserving surgical sensitivity while increasing resource use through reduced specificity. We evaluated this safety-stewardship trade-off across frontier large language models (LLMs) to provide an exploratory diagnostic benchmark.MethodsWe performed a cross-sectional in-silico study using 50 standardized synthetic vignettes (25 urgent surgical, 25 non-urgent) internally developed by our investigative group. GPT-5.5, Claude 4.0 Opus, and Gemini 3.0 Ultra were evaluated using identical zero-shot deterministic prompting against an internally established, blinded dual-specialty expert consensus.ResultsExact primary-diagnosis accuracy was 92.0% (95% CI, 81.2-96.8%) for GPT-5.5, 88.0% (95% CI, 76.2-94.4%) for Claude 4.0 Opus, and 76.0% (95% CI, 62.6-85.7%) for Gemini 3.0 Ultra. Surgical sensitivity remained high across models (84.0% to 96.0%), whereas specificity was lower (64.0% to 80.0%) than specialist consensus. Models demonstrated a conservative escalation pattern with frequent over-triage decisions. Positive likelihood ratios ranged from 2.33 to 4.60. Over-investigation bias was the predominant error pattern.ConclusionIn this exploratory benchmark, frontier LLMs consistently prioritized surgical sensitivity over specificity in internally authored structured acute-abdomen vignettes. These findings reflect differences in performance on controlled synthetic scenarios rather than independently validated clinical diagnostic accuracy. Claims regarding clinical patient safety or readiness for implementation require future prospective validation with independent, real-world clinical cases and external multidisciplinary adjudication.

The American Surgeon
Ümraniye Eğitim ve Araştırma Hastanesi (TR), Sağlık Bilimleri Üniversitesi (TR)
Good health and well-being
Artificial Intelligence in Healthcare and Education
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon
When Artificial Intelligence Plays It Safe: Large Language Models Preserve Surgical Sensitivity at the Cost of Diagnostic Specificity in Acute Abdominal Pain — Fatih Başak, Sıbel Ocak Serın · The American Surgeon (2026) | TGRS Research Map | TGRS