Multi-endpoint AI morphology model (MEAM) enhances risk prediction for vascular events and disease progression in MPNs

Abstract

Abstract INTRODUCTION The myeloproliferative neoplasms (MPNs) are clonal haematological disorders that display significant clinical heterogeneity. A significant challenge remains identifying patients at increased risk of death, life-altering vascular events or disease transformation (secondary myelofibrosis or blast phase disease/AML). Conventional risk predictors incorporate patient age, blood parameters, molecular profiling and previous thrombosis. However, these risk predictors are restricted to specific clinical events, incompletely capture the dynamic nature of MPN and have limited scope to detect therapy-driven disease modification. Improved dynamic personalised risk prediction that captures fundamental aspects of bone marrow (BM) health in MPN have potential to complement and enhance existing risk models and support the evaluation of new treatments. Here, we propose a novel multi-endpoint AI morphological model (MEAM) that derives risk scores from Essential Thrombocythemia (ET) and Polycythemia Vera (PV) BM trephine (BMT) samples. The adverse events captured include vascular events (thrombosis and haemorrhage), fibrotic (MF) transformation, blast phase / AML transformation and overall survival (OS). METHODS Our training data consists of 949 patients drawn from independent MPN cohorts [PT1 (n=545), MAJIC (n=216), and supplementary internal sources (n=188)], including 643 ET, 145 PV, 39 MF, 102 normal, and 20 AML cases. A total of 1661 H&E WSIs were used. Using a Cox proportional hazards (CPH) framework, we trained a vision transformer to predict multi-dimensional log hazard ratios for four survival endpoints, with built-in attention heatmaps for interpretability. Patch-level features were extracted at 20x magnification using the CTransPath foundation model. The algorithm was developed using 5-fold cross-validation, achieving an average C-index of 0.77 ± 0.017. Predictions on the screening samples from the validation folds were used to illustrate the tool's clinical utility. RESULTS In ET patients, MEAM matched or exceeded established clinical models in C-index: OS prediction (0.66 vs IPSET = 0.63) and vascular events (0.60 vs revised IPSET thrombosis R-IPSET-T = 0.64). Although no validated risk models currently exist to predict transformation in chronic MPN, we also compared our model's performance against IPSET. MEAM outperformed secondary MF prediction (0.68 vs IPSET =0.57) and AML transformation (0.76 vs IPSET =0.52). Combining MEAM with conventional risk scores markedly improved the prediction across all endpoints: OS [MEAM+IPSET=0.71 (+13%)], vascular events [MEAM+R-IPSET-T=0.68 (+5.9%)], MF transformation [MEAM+IPSET=0.68 (+19%)] and AML transformation [MEAM+IPSET=0.75 (+44%)]. In PV patients, MEAM improved risk prediction over conventional models that incorporates age and previous thrombosis: C-index in progression to secondary MF [MEAM=0.67, conventional=0.50, combined=0.66 (+31%)] and AML progression [MEAM=0.69, conventional=0.56, combined=0.70 (+24%)]. MEAM brought an uplift of 0.02% and 4.1% to vascular and death risk predictions, respectively to the conventional C-index (0.53 and 0.62) when used in conjunction. With an optimised threshold for time-dependent AUROC, ET patients were binarized into low/high risk groups. MEAM stratified all four endpoints significantly (log-rank p < 0.01). Within 5 years of biopsy 80 ET patients experienced a vascular event, with R-IPSET-T failing to flag 25 as being high risk. Notably, MEAM flagged 10 of these cases correctly as high risk. Of 35 patients younger than 60 at diagnosis in whom there was a subsequent vascular event, R-IPSET-T misclassified 77% (27/35) as (very) low/intermediate risk. MEAM correctly identified 10/27 (37%) of these as high risk. Among 34 deaths in this younger age group, IPSET labelled all as low/intermediate risk. By contrast, MEAM identified 13/34 (38%) as high risk. CONCLUSIONS We present an AI-based morphological model that enhances risk stratification in ET and PV. Our morphology predictor offers prediction power comparable or superior to established clinical risk predictors for vascular events, disease progression and OS. Using MEAM to augment existing and newly emerging prognostic tools, we envision an interpretable and dynamic morphological risk score that is ideally suited to routine clinical application and can support the development of novel therapies that aim to induce meaningful disease modification in MPN.