Automatic prediction of patients early discontinuation in oncology clinical trials.

BACKGROUND: Early discontinuation (ED) in clinical trials (CTs) is frequent and deleterious for the patients, the care team, and the study duration. ED comprises screening failure or discontinuation during the first month of the treatment phase, and is often difficult to predict by clinicians. We aim at predicting ED by automatic analysis of patient's clinical record using language models (LMs). MATERIALS AND METHODS: We fine-tuned a French LM on the oncology clinical reports of a French cancer center, and obtained a pretrained LM named OncoBERT. We then selected consultation reports of patients included in oncology CTs for any tumor type that we used to fine-tune OncoBERT and obtained a new model for ED prediction. We carried out a retrospective and prospective evaluation and used eXplainable Artificial Intelligence (XAI) methods to interpret the predictions. RESULTS: On the retrospective test cohort of 1007 reports, the model achieved a precision of 0.77, recall of 0.95, and could have decreased the ED rate from 25.3% to 21.3%. On the prospective test cohort it reached a precision of 0.75, recall of 0.75, and could have decreased the ED rate from 33.7% to 27%. Using XAI showed that the words used by the model to predict ED reflect deterioration of general condition, a well-known factor of ED. CONCLUSION: We have developed a LM that is portable, explainable, with near-human performances for ED prediction in oncology CTs. We anticipate that democratization of automatic trial matching tools should be complemented by ED prediction tools to fully optimize access to CTs and patient recruitment.