🇪🇸 Leer este artículo en español

The patient who drops out

A patient in an oncology clinical trial may leave the study prematurely. We are not talking about disease progression: we are referring to early discontinuation (ED), which includes screening failure or dropout during the first month of treatment. It is common, harms the patient, strains the team and prolongs the trial. And despite its impact, clinicians find it hard to predict.

What if an AI could read the medical record and warn us before it happens?

What they did: a BERT for French clinical reports

A team from the Centre Léon Bérard (Lyon) developed OncoBERT, a language model pre-trained in French and specifically fine-tuned on oncology clinical reports from a French cancer centre. They then further refined it for the specific task: given a consultation report of a patient enrolled in a clinical trial (of any tumour type), predict whether ED will occur.

Evaluation was twofold:

	Retrospective	Prospective
Reports	1,007	—
Precision	0.77	0.75
Sensitivity (recall)	0.95	0.75
ED rate (actual → potential)	25.3% → 21.3%	33.7% → 27%

In the retrospective cohort, if the model had flagged high-risk patients, the ED rate could have dropped from 25.3% to 21.3%. In the prospective cohort, from 33.7% to 27%.

The most interesting part: what the model looks at

Using explainable AI (XAI) methods, the authors checked which words drove the predictions. The result is reassuring: the model focuses on expressions that reflect decline in performance status — the same factor clinicians know is associated with discontinuation.

In other words, it’s not a black box “guessing” from rare signals: it learns to read what is already written in the record, and does so consistently and at scale.

Critical reading

• Strength: design with both retrospective and prospective validation, using real data from an oncology centre. The model is portable and explainable.
• Limitations: single-centre study, French language, reports from a specific system — generalisation to other languages and hospital settings remains to be proven.
• Important nuance: sensitivity drops from 0.95 (retrospective) to 0.75 (prospective). The “almost human” performance claimed in the conclusion is not quantified against a direct comparison with clinicians using abstract data.
• Interpretation: the model does not replace the research team; it alerts so that the team can intervene (psychosocial support, follow-up adjustment) before dropout occurs.

💡 Takeaway

Automatic prediction of early dropout in clinical trials is technically viable, portable and — crucial for clinical practice — explainable: the model uses the same signs of deterioration we use, but systematically. Such tools could complement automatic trial matching to optimise access to trials and care for participating patients.

📄 Paper: Automatic prediction of patients early discontinuation in oncology clinical trials — ESMO Real World Data and Digital Oncology, 2026.

— This analysis was generated by ANGIE (Always Next to Guide, Inspire and Empower), an artificial intelligence system with SOUL profiles, designed by Dr. Javier Pumares Pérez.

Connect on LinkedIn · ppj.es