Non-invasive differentiation between aplastic anemia and myelodysplastic syndromes based on an 8-feature lightGBM model: A multicenter external validation study.
Source: PubMed, NCBI / U.S. National Library of Medicine
Given the highly overlapping pancytopenic phenotypes, non-invasive differentiation between aplastic anemia (AA) and myelodysplastic syndromes (MDS) remains a formidable clinical challenge. The current diagnostic gold standard relies on invasive bone marrow biopsy and lacks objective standardization. Here, we develop and externally validate a robust, interpretable machine-learning framework utilizing routine peripheral blood parameters to optimize clinical triage. We retrospectively enrolled patients with histopathologically confirmed AA or MDS, partitioning them into a training set (n = 310) and an internal validation set (n = 131). An independent external cohort (n = 74) was leveraged to evaluate cross-institutional generalizability. Following a rigorous multi-algorithm feature selection framework (incorporating XGBoost, Random Forest, and SVM-RFE), we systematically evaluated 14 machine-learning classifiers. We applied the SHapley Additive exPlanations (SHAP) framework to decode pathophysiological drivers and constructed a visual clinical nomogram for point-of-care application. Algorithmic intersection distilled the high-dimensional data into a parsimonious 8-feature panel (CHOL, hsCRP, IL-6, SAA, AGE, LDL-C, HRF, and IL-10). The optimized LightGBM model demonstrated superior discriminative accuracy. Crucially, the model exhibited robust performance in the independent external validation cohort, achieving an area under the curve (AUC) of 0.997 and an overall accuracy of 97.
Abstract
Given the highly overlapping pancytopenic phenotypes, non-invasive differentiation between aplastic anemia (AA) and myelodysplastic syndromes (MDS) remains a formidable clinical challenge. The current diagnostic gold standard relies on invasive bone marrow biopsy and lacks objective standardization. Here, we develop and externally validate a robust, interpretable machine-learning framework utilizing routine peripheral blood parameters to optimize clinical triage. We retrospectively enrolled patients with histopathologically confirmed AA or MDS, partitioning them into a training set (n = 310) and an internal validation set (n = 131). An independent external cohort (n = 74) was leveraged to evaluate cross-institutional generalizability. Following a rigorous multi-algorithm feature selection framework (incorporating XGBoost, Random Forest, and SVM-RFE), we systematically evaluated 14 machine-learning classifiers. We applied the SHapley Additive exPlanations (SHAP) framework to decode pathophysiological drivers and constructed a visual clinical nomogram for point-of-care application. Algorithmic intersection distilled the high-dimensional data into a parsimonious 8-feature panel (CHOL, hsCRP, IL-6, SAA, AGE, LDL-C, HRF, and IL-10). The optimized LightGBM model demonstrated superior discriminative accuracy. Crucially, the model exhibited robust performance in the independent external validation cohort, achieving an area under the curve (AUC) of 0.997 and an overall accuracy of 97.30%. SHAP analysis bridged mathematical predictions with pathophysiology, revealing that advanced age and an "inflammaging" cytokine profile (IL-6, hsCRP, SAA) are the strongest predictive features associated with MDS. Conversely, distinct lipid remodeling (CHOL, LDL-C) and erythropoietic alterations (HRF) shifted the diagnostic probability toward AA. In conclusion, we established and externally validated a highly accurate, non-invasive 8-feature LightGBM diagnostic model. Translated into a practical clinical nomogram, this interpretable artificial-intelligence tool serves as an efficient "triage gatekeeper" to minimize unnecessary invasive biopsies, thereby facilitating precision triage in hematology.