PIDgeon: An Explainable AI Model for Improved Flow Cytometry-Based Screening of Lymphoid Primary Immunodeficiencies.

Area of Science:

• Computational immunology
• Bioinformatics
• Artificial intelligence in healthcare

Background:

• Primary immunodeficiencies (PIDs) are rare immune system disorders often diagnosed using multi-parameter flow cytometry (FCM).
• Manual analysis of FCM data is subjective and time-consuming.
• An automated computational pipeline is needed to improve PID diagnostics.

Purpose of the Study:

• To present PIDgeon, a fully automated AI-based pipeline for analyzing FCM data.
• To characterize PID immune profiles and suggest PID subtypes.
• To generate interpretable reports for clinical use.

Main Methods:

• The PIDgeon pipeline utilizes FlowSOM and extreme gradient boosting models.
• It was trained and validated on standardized FCM data from healthy controls and PID patients.
• Multi-centric validation was performed on independent internal and external datasets.

Main Results:

• PIDgeon achieved high accuracy (correlation > 0.90) in cell count enumeration for lymphocyte subsets.
• It demonstrated high sensitivity (93-100%) in predicting severe T-cell PIDs.
• Low false-negative rates (1.5-5.4%) were observed for distinguishing PID from non-PID controls, with initial subtype prediction capabilities.

Conclusions:

• PIDgeon is an accessible and explainable AI pipeline for early PID diagnostics.
• It enhances data analysis efficiency for laboratory immunologists.
• The tool aligns with current clinical needs in PID diagnosis.