Development and validation of a machine learning-based diagnostic model for identifying nonneutropenic invasive...

This study aims to develop and validate an optimized diagnostic model for nonneutropenic invasive pulmonary aspergillosis (IPA) among suspected cases. A cohort of 344 nonneutropenic suspected cases from 13 medical centers (August 2020 to February 2024) was analyzed. The cohort was divided into a training data set (70%) and a testing data set (30%) using stratified sampling based on the IPA diagnosis. Three machine learning models (a regularized logistic regression model, a support vector machine model, and a weighted ensemble model) were developed. SHapley Additive explanation (SHAP) method was used for model interpretation. Six predictor variables were finally selected: sputum Aspergillus culture, Aspergillus-specific IgG, imaging feature of cavity, serum galactomannan, critical condition, and plasma pentraxin 3. The weighted ensemble model, exhibiting the significantly higher specificity of 95.1% in internal cross-validation and 95.7% in testing among the three models, was selected as the optimal prediction model despite comparable discrimination capacity, calibration ability, and clinical applicability across all models. The risk score derived from SHAP values showed a highly significant correlation with the predicted probability of the weighted ensemble model (Spearman ρ = 0.974), achieving an area under the curve of 0.857 in internal cross-validation and 0.871 in external testing. Using the optimal cut-off value of 3, the risk score demonstrated sensitivity (68.8%) and specificity (87.5%) comparable to those of bronchoalveolar lavage fluid galactomannan (cut-off = 1.0). The diagnostic model and risk score could assist in identifying nonneutropenic IPA from suspected cases independently of invasive procedures, thereby enhancing clinical applicability.

Although clinicians can screen out suspected cases through medical history inquiries, the diagnosis of nonneutropenic invasive pulmonary aspergillosis (IPA) from suspected cases remains a significant challenge. The study developed a novel diagnostic framework by integrating clinical parameters, imaging features, and laboratory biomarkers using machine learning techniques. The risk score, derived from SHapley Additive explanation values, exhibited a highly significant correlation with the predicted probability of the weighted ensemble model, demonstrating robust discrimination capacity and generalizability. The diagnostic model and risk score could assist in identifying nonneutropenic IPA from suspected cases independently of invasive procedures, thereby enhancing clinical applicability.