Celfib-AI: A machine learning model to predict bone marrow fibrosis and cellularity from complete blood count – a multicenter model development and validation study in France

Abstract

Abstract Introduction Bone marrow biopsy is an essential component of the WHO diagnostic criteria for myeloproliferative neoplasms (MPNs). However, it remains an invasive procedure that can be difficult to perform in elderly or anticoagulated patients. This study aimed to develop an artificial intelligence (AI) model to predict two key and reproducible histopathological findings from bone marrow biopsy—grade 2–3 fibrosis and overall cellularity—using complete blood count (CBC) data in patients with suspected or confirmed MPNs. Methods Histopathological and CBC data from 1,691 bone marrow biopsies performed at diagnosis or during follow-up were retrospectively collected across 10 French university hospitals (Lyon n=761; Brest n=214; Grenoble n=154; Clermont-Ferrand n=127; Nancy n=127; Paris–Mondor n=102; Saint-Étienne n=71; Bordeaux n=61; Paris–Kremlin-Bicêtre (Paris-KB) n=48; Nantes n=26). Thirty-three CBC variables and patient age at biopsy were used as predictors. For each case, histological findings were controlled by a referent pathologist. Equivocal cases underwent central consensus review by an expert panel (GEBOM group). Data from six centers (Lyon, Grenoble, Paris–KB, Clermont-Ferrand, Nantes, Bordeaux) were used for model development and split into training (75%) and internal test (25%) cohorts, stratified by center and fibrosis grade. The remaining four centers were used for external validation. After model selection, an additional prospective “real-life” series of 144 consecutive biopsies performed for any indication in Lyon was evaluated. Linear (GLM, Lasso, Elastic-Net), support-vector machine (SVM), tree-based (XGboost, Random Forest), Bayesian (Naïve Bayes), neural network (neural net) and tabular foundation (tabFPN) models were evaluated to predict bone marrow fibrosis (binary model: no significant fibrosis vs. grade 2–3 fibrosis) and cellularity (multiclass model: decreased, normal, increased), using 10-fold cross-validation on the training set. The best model was selected based on its performance on the internal test set and further evaluated on the external validation sets and the prospective cohort. Results Among the 1,691 evaluated bone marrow biopsies, histopathological diagnosis was essential thrombocythemia (n=625, 37.0%), not MPN (n=380, 22.5%), primary myelofibrosis (PMF, n=220, 13.0%), polycythemia vera (n=186, 11.0%), secondary myelofibrosis (n=126, 7.5%), pre-PMF (n=81, 4.8%), and unclassifiable MPN or MDS/MPN overlap (n=72, 4.3%). Among MPN patients, the main driver mutations were JAK2V617F (n=886, 67.6%), CALR (n=205, 15.6%), MPL (n=51, 3.9%) and triple negative status (n=129, 9.8%). The XGBoost binary model achieved the best performance for fibrosis prediction, with an AUROC [95% CI] of 0.96 [0.95–0.97] in the training set and 0.91 [0.86–0.95] in the internal test set. Performance was confirmed in validation cohorts with complete CBC data (Paris–Mondor (AUROC 0.94 [0.89-1.00]) and Nancy (0.98 [0.96–1.00])), and remained robust even with partially missing data (Brest, lacking erythroblast count, circulating myeloid precursors, IDR, and MPV (AUROC 0.84 [0.76–0.92]), and Saint-Étienne, missing MPV and IDR (AUROC 0.83 [0.74–0.93])). Prospective evaluation in real-life setting revealed an AUROC of 0.96 [0.92-0.99] and an accuracy of 88% [81.9-92.4] in the Lyon-prospective cohort. Sensitivity analyses showed stable model performance with as few as 10 variables, including (in order of importance): circulating myeloid precursors percentage and count, erythroblast percentage, hemoglobin level, IDR, MPV, white blood cells count, monocyte percentage, age and MCH. For cellularity prediction, the multiclass XGboost model demonstrated the highest discriminatory performance, with AUROC [95%CI] for decreased / normal / increased cellularity of 0.90 [0.89-0.92] / 0.86 [0.84-0.89] / 0.89 [0.87-0.91] in the training set and 0.72 [0.63-0.8] / 0.76 [0.7-0.80] / 0.81 [0.76-0.86] in testing set, respectively. Conclusion These findings indicate that grade 2–3 bone marrow fibrosis can be accurately predicted using routine CBC parameters and an XGBoost model in patients with suspected or confirmed MPNs. Predictive modeling of bone marrow cellularity yielded lower yet interesting accuracy. A web-based tool designed for routine clinical use is currently under development under evaluation in a prospective setting across 4 independent centers.