# Predicting modelling for early diagnosis in early myeloproliferative neoplasms, not otherwise specified. Evidence from a machine learning study

## TL;DR

This study uses machine learning to help diagnose early myeloproliferative neoplasms by analyzing bone marrow biopsy images, showing promising results in supporting clinical decisions.

## Contribution

The study introduces a machine learning model for classifying early myeloproliferative neoplasms using morphological features from bone marrow biopsies.

## Key findings

- The model achieved a mean multiclass AUC of 0.70 and accuracy of 0.60 in classifying MPN subtypes.
- On MPN-NOS patients, the model agreed with pathologists 61.5% of the time and with clinical follow-up 50% of the time.

## Abstract

Recently, artificial intelligence (AI) has become a valuable tool for diagnosing and predicting outcomes in blood disorders. Whole Slide Imaging (WSI) of bone marrow biopsies (BMBs) offers detailed, high-resolution views of cells and tissues; its adoption may improve the resources dedicated to the interpretation of BMB with suspected early myeloproliferative neoplasms (MPNs).

We collected a retrospective dataset of H&E-stained BMBs from 88 patients diagnosed with MPN, divided into three groups: 19 with prefibrotic primary myelofibrosis (pre-PMF), 30 with polycythemia vera (PV), and 39 with essential thrombocythemia (ET). Using AI, we framed this as a three-class classification problem. For each whole slide image, we automatically calculated cellularity and cell density. We extracted morphological features related to megakaryocytes—area, perimeter, and circularity—and summarized them for each patient using statistics like mean, standard deviation, skewness, kurtosis, and entropy. This resulted in 17 features combining nuclear morphology, cell density statistics, and cellularity. After selecting significant features with the Kruskal-Wallis test, we trained a Support Vector Machine (SVM) classifier with 5-fold cross-validation to predict MPN subtypes. We then tested the model on 13 patients with a diagnosis of MPN not otherwise specified (MPN-NOS) to assess its capacity to correctly characterize the diagnosis.

The model achieved a mean multiclass Area Under the Curve (AUC) of 0.70 ± 0.01 and an accuracy of 0.60 ± 0.03. When tested on the 13 patients with MPN not otherwise specified (MPN-NOS), the model agreed with pathologists’ biopsy classifications 61.5% of the time and with clinical follow-up evaluations 50% of the time.

This study represents a first step toward the development of automated tools to support MPN diagnosis, providing potential assistance to haematologists and pathologists in the clinical management of patients.