Abstract
BACKGROUND: Thalassaemia is one of the most common inherited monogenic diseases worldwide with a heavy global health burden. Considering its high prevalence in low and middle-income countries, a cheap, accurate and high-throughput screening test of thalassaemia prior to a more expensive confirmatory diagnostic test is urgently needed. METHODS: In this study, we constructed a machine learning model based on MALDI-TOF mass spectrometry quantification of haemoglobin chains in blood, and for the first time, evaluated its diagnostic efficacy in 674 thalassaemia (including both asymptomatic carriers and symptomatic patients) and control samples collected in three hospitals. Parameters related to haemoglobin imbalance (α-globin, β-globin, γ-globin, α/β and α-β) were used for feature selection before classification model construction with 8 machine learning methods in cohort 1 and further model efficiency validation in cohort 2. RESULTS: The logistic regression model with 5 haemoglobin peak features achieved good classification performance in validation cohort 2 (AUC 0.99, 95% CI 0.98-1, sensitivity 98.7%, specificity 95.5%). Furthermore, the logistic regression model with 6 haemoglobin peak features was also constructed to specifically identify β-thalassaemia (AUC 0.94, 95% CI 0.91-0.97, sensitivity 96.5%, specificity 87.8% in validation cohort 2). CONCLUSIONS: For the first time, we constructed an inexpensive, accurate and high-throughput classification model based on MALDI-TOF mass spectrometry quantification of haemoglobin chains and demonstrated its great potential in rapid screening of thalassaemia in large populations.Key messagesThalassaemia is one of the most common inherited monogenic diseases worldwide with a heavy global health burden.We constructed a machine learning model based on MALDI-TOF mass spectrometry quantification of haemoglobin chains to screen for thalassaemia.