Artificial intelligence-based screening of paroxysmal nocturnal hemoglobinuria clones: Development and comparative evaluation of machine learning models

The diagnosis of paroxysmal nocturnal hemoglobinuria (PNH) relies on detecting cells deficient in GPI-anchored proteins and is currently based on manual, time-consuming interpretation requiring specialized expertise. This study aimed to develop machine learning models to identify PNH clones in granulocytes and monocytes using flow cytometry data containing information on cell size, complexity, and expression of FLAER, CD15, CD45, CD64, and CD157. For training, 20 normal and 20 PNH samples were randomly selected; normal samples were used in full, whereas only PNH cells were retained from PNH samples. For analytical validation, 22 normal and 18 PNH samples were analyzed in their entirety. Samples were processed through a preprocessing pipeline and split into training (12 normal, 12 PNH), validation (4 normal, 4 PNH), and test (4 normal, 4 PNH) datasets. Two supervised models were developed: a Random Forest (RF) and an Artificial Neural Network (ANN). Validation samples underwent the same preprocessing, except for concatenation. Both models achieved specificity close to 100%, accuracy >93%, ROC-AUC >0.99, and PR-AUC >0.98; the ANN showed higher sensitivity (0.85 vs. 0.78) and F1-score (0.92 vs. 0.87). In qualitative analysis, RF and ANN achieved 100% and 45% agreement with manual analysis, respectively, and quantitative analysis confirmed statistical equivalence between RF and manual results. RF repeatability was confirmed by identical triplicate results in 10 samples. In spiking experiments, RF identified >50 PNH cells in all samples spiked with 200 or 100 cells, in 3 of 5 samples spiked with 50 cells, and in none spiked with 25 cells. These results indicate that the RF model shows promise for PNH analysis; however, clinical implementation will require integration into immunophenotyping software and post-deployment performance monitoring.