Jean-Emmanuel Bibault, MD, PhD (@jebibault) header
Jean-Emmanuel Bibault, MD, PhD (@jebibault) profile picture

Professor of Radiation Oncology @APHP @univ_paris_cite | AI researcher @INSERM | Author

Paris
Joined June 2026
1/3 Can a mammogram flag breast cancer 4–6 years before diagnosis? A Norwegian retrospective cohort tested a commercial detection AI in 116,495 women across three biennial screening rounds.
3
1
24
2/3 AI scores were higher in the breast that later developed cancer. For screening-detected cancers, AUC rose from 0.64 at the earliest round to 0.97 at the diagnostic round; for interval cancers, it rose from 0.66 to 0.78.
1
12
1/3 An AI can analyze 39 features of an echocardiogram. How should we use it? PanEcho was developed and validated using 1.2 million videos from 32,265 examinations in 24,405 patients, with additional external validation.
1
10
2/3 It performed 18 classification tasks with a median internal AUROC of 0.91, plus 21 measurement tasks. For left ventricular ejection fraction, mean absolute error was 4.2 percentage points internally and 4.5 externally.
1
7
1/4 A doctor’s final approval can still follow an AI-induced error. In a randomized experiment, 225 physicians assessed neurocritical-care cases with the same AI instructed to support correct decisions or covertly steer them toward harmful ones.
1
4
3/4 This does NOT estimate the harm rate of ordinary clinical AI: the adversarial model was deliberately instructed to mislead. It tests whether medical expertise alone resists that manipulation. Experience offered no clear protection in this experiment.
1
1
4/4 Evaluate how AI advice changes clinical decisions, including initially correct ones. A final human signature is an incomplete safety measure. Preprint, not peer-reviewed: doi.org/10.64898/2026.10.05.…
1/3 Ten million ECGs. What does that actually prove? ECGFounder was built using 10.77 million recordings from 1.82 million people, covering 150 diagnostic labels. It exceeded an AUROC of 0.95 for 80 diagnoses in internal validation.
1
4
2/3 External testing and adaptation to additional tasks make this more useful than a single-purpose classifier. But 10 million recordings aren't 10 million independent patients. And AUROC doesn't tell you how many false alerts a clinic will receive.
1
4
1/4 Can radioligand therapy extend disease control before SBRT? LUNAR randomized 92 men with hormone-sensitive oligorecurrent prostate cancer and 1–5 PSMA-PET lesions to SBRT alone or two cycles of 177Lu-PNT2002 followed by SBRT.
1
8
3/4 No new safety signal emerged. Limits: single-center, open-label phase II; 87 analyzed after five control-arm withdrawals. Overall survival benefit is unproven. Conventional-imaging PFS was exploratory and not statistically significant.
1
4/4 A strong rationale for confirmatory trials. Does delaying PET-defined progression yield durable benefits that justify adding radioligand therapy to SBRT? That remains the clinical question. doi.org/10.1016/j.eururo.202…
1/3 Can a scan of your palm help detect sepsis? In a prospective ICU study, AI analyzed hyperspectral skin images. Among 437 patients with an established sepsis status, imaging alone achieved an AUROC of 0.80.
1
2
2/3 Adding clinical and laboratory data raised the AUROC to 0.94. That distinction matters: the headline performance came from combining information. A specialized camera also isn't a smartphone photo.
1
1