AI-based prescreening of clonal hematopoiesis in patients with liver disease using Sysmex XN hematology analyzer data

Clonal hematopoiesis (CH) is characterized by the clonal expansion of blood cells harboring somatic mutations and is commonly observed with aging. Although CH is known to induce hemogram alterations, differentiating it using standard complete blood count (CBC) metrics alone is challenging, highlighting the need for multidimensional hematologic profiling. To address this, this study leverages expanded CBC parameters from the Sysmex XN analyzer, incorporating broader optical measurements, to explore whether they improve CH detection. Given prior evidence linking CH to liver pathology, this study focuses on the relationship between CH and liver disease. Associations between the two have been demonstrated in large-scale cohorts such as the UK Biobank, and CH-driven macrophage activation has been implicated in the progression of NAFLD, NASH, cirrhosis, and hepatocellular carcinoma. However, liver diseases-including cirrhosis-produce characteristic CBC alterations, making CH discrimination particularly difficult. This motivates evaluating whether an AI model can detect CH-related signals despite liver disease-induced hematologic changes. At Seoul National University Hospital, with support from Sysmex, raw outputs from XN analyzers have been prospectively archived since 2022, covering 146 parameters, including reportable items and research indices. Using these data, 2,173 CBCs from 303 patients with liver diseases were analyzed using diverse machine- and deep-learning methods. Initial models were trained on the first test data per patient. To capture the longitudinal dynamics of CH, features were reconstructed from all repeated tests per patient (mean, standard deviation, maximum, minimum), and a soft-voting ensemble combining logistic regression, LightGBM, and CatBoost was developed, yielding stable performance. A deep-learning MIL model was applied to aggregate repeated CBC measurements across patient time points and capture instance-level contributions, complemented by an autoencoder to compress correlated features into latent clusters and improve model stability. With first-visit data only, logistic regression achieved precision 41.0%, recall 51.6%, F1 0.457; CatBoost achieved precision 39.2%, recall 93.5%, F1 0.552. When summary statistics with soft voting model were used, performance improved to precision 72.2%, recall 80.0%, F1 0.759. Within the MIL framework, analysis of model-weighted instances revealed distinct activation patterns in autoencoder-derived latent clusters associated with liver diseases, offering a novel interpretability perspective beyond traditional feature-importance analysis. This algorithm provides practical guidance to identify CH from CBC data in patients with liver disease, enabling prescreening prior to precision diagnostics and supporting timely diagnosis and improved management.