Prediction of bone marrow fibrosis from complete blood count in myeloproliferative neoplasms (FIBOM-AI): a...

Bone marrow biopsy is essential for myeloproliferative neoplasms diagnosis, particularly for assessing bone marrow fibrosis. However, it is an invasive procedure that can be difficult to perform in some patients. This study aimed to develop an artificial intelligence (AI) model (FIBOM-AI) to predict grade 2-3 fibrosis using complete blood count (CBC) data and age.

In this machine learning model, histopathological and CBC data from patients (no age restriction) undergoing bone marrow biopsies, performed either at diagnosis or during follow-up for suspected or established myeloproliferative neoplasm, were retrospectively collected across 13 French university hospitals. 27 CBC variables and patient age at biopsy were used as predictors. Data from six independent centres were used for model development and the remaining seven centres were used for external validation. The primary outcome was grade 2-3 fibrosis as a binary endpoint. Nine algorithms were trained using 10-fold cross validation. The final model was evaluated in a real-life prospective setting on all bone marrow biopsies performed in four French centres and one Canadian centre. To guide clinical decisions, two different predictions based on the final model were set: an overall prediction, maximising the overall accuracy of the model, and a confident prediction, maximising either sensitivity (rule-out) or specificity (rule-in).

Between Jan 1, 2014, and Dec 31, 2023, 1995 patients from 13 centres were included in the retrospective cohort; the median age was 62·0 years (IQR 50·0-71·0), 1062 (53·2%) were male, and 933 (46·8%) were female. Between Jan 1, 2024, and Dec 31, 2024, 493 patients from five centres were included in the prospective cohort; the median age was 65·0 years (IQR 53·6-73·0), 297 (60·2%) were male, and 196 (39·8%) were female. The Extreme Gradient Boosting model had the best performance, with an area under the curve of 0·96 (95% CI 0·95-0·97), 0·90 (0·85-0·95), and 0·92 (0·90-0·95) on the training, testing, and validation sets, respectively. The final accuracy was 88·1%, 87·0%, and 88·6% for the overall predictions and 99·5%, 94·5%, and 96·9% for the confident predictions in the training, testing, and validation sets, respectively. Prospective evaluation showed an accuracy of 85·2% (range 82·0-87·3) for the overall predictions and 98·6% (96·0-100·0) for the confident predictions.

FIBOM-AI can be reliably used as a triage and prioritisation tool to support timely bone marrow evaluation or provide complementary risk assessment when bone marrow biopsy is not immediately feasible.

None.