Optimising criteria for manual smear review following automated blood count analysis: A machine learning approach
Abstract
The complete blood count (CBC) performed by automated haematology analysers is the most common clinical procedure in the world. Used for health checkup, diagnosis and patient follow-up, the CBC impacts the majority of medical decisions. If the analysis does not fit an expected setting, the laboratory staff manually reviews a blood smear, which is highly time-consuming. Criteria for reviewing CBCs are based on international consensus guidelines and locally adjusted to account for laboratory resources and populations characteristics. Our objective is to provide a clinical laboratory decision support tool to identify which CBC variables are linked to an increased risk of abnormal manual smear and at which threshold values. Thus, we treat criteria adjustment as a feature selection problem. We propose a cost-sensitive Lasso-penalised additive logistic regression combined with stability selection, adapted to the peculiarities of data and context: class-imbalance, categorisation of continuous predictors, required stability and enhanced interpretability. Using simulated and real CBC data, we show that our proposal is competitive in terms of predictive performance (compared to deep neural networks) and model selection performance (provided that there is sufficient data in the neighbourhood of the true thresholds). The R code is publicly available as an open source project.
Domains
Statistics [stat] Machine Learning [stat.ML] Statistics [stat] Methodology [stat.ME] Statistics [stat] Computation [stat.CO] Statistics [stat] Applications [stat.AP] Computer Science [cs] Machine Learning [cs.LG] Life Sciences [q-bio] Santé publique et épidémiologie Life Sciences [q-bio] Human health and pathology Hematology
Origin : Files produced by the author(s)