Machine Learning Models Improve the Diagnostic Yield of Peripheral Blood Flow Cytometry

Area of Science:

• Hematology
• Machine Learning
• Medical Diagnostics

Background:

• Peripheral blood flow cytometry (PBFC) aids in evaluating circulating hematologic malignancies (HM).
• However, PBFC has limitations in screening utility.
• Machine learning (ML) may enhance PBFC's diagnostic value by integrating clinical history and complete blood count (CBC) data.

Purpose of the Study:

• To assess if ML models incorporating clinical history and CBC/differential parameters can improve the utilization of PBFC.
• To develop predictive models for PBFC results.

Main Methods:

• PBFC cases with concurrent CBC/differential data were divided into training (n=626) and testing (n=159) sets.
• PBFC results with abnormal blast/lymphoid populations were classified as positive.
• Two ML models (decision tree and logistic regression) were used to predict PBFC results.

Main Results:

• Positive PBFC results were significantly more frequent in patients with prior HM (58% vs. 21%).
• Neutrophil percentage, absolute lymphocyte count, and blast percentage were key predictors (AUC > 0.7).
• The decision tree model showed 98% sensitivity and 65% specificity (AUC=0.906), while logistic regression achieved 100% sensitivity and 54% specificity (AUC=0.919).

Conclusions:

• ML-based triaging strategies can effectively decrease unnecessary PBFC utilization.
• An estimated reduction of 35% to 40% in unnecessary tests is achievable.
• This approach optimizes diagnostic workflows for hematologic malignancies.