Abstract Study Objectives Differential diagnosis of narcolepsy type 2 (NT2) from type 1 (NT1) and idiopathic hypersomnia (IH) is challenging due to overlapping symptoms. We developed an automated method using nocturnal polysomnography (nPSG) data to differentiate these conditions and clinical controls (CCs), and explored varying sleep phenotypes within NT1, NT2, IH, and CCs. Methods We analyzed nPSG data from drug-free individuals with NT1, NT2, and IH, or CCs. Sleep features were derived at whole-night and per-quarter-night levels, including hypnogram, transition probability, hypnodensity, spindle, and quantitative electroencephalogram (qEEG) features. Random forest machine learning models were used for three classification tasks. Within-diagnosis clustering identified potential diagnosis subgroups. Results The sample included 350 individuals (52% females; median age 30 years; 114 NT1, 90 NT2, 105 IH, 41 CCs). Our models achieved area under the receiver operating characteristic curve values of 0.87, 0.79, and 0.82 for distinguishing NT2 from CCs, NT2 from IH, and IH from CCs, with corresponding F1 scores of 0.74, 0.71, and 0.69, respectively. qEEG features substantially contributed to model performance distinguishing NT2 from IH. Cluster analysis revealed two NT1 subgroups (one showing more severe sleep disturbances), two NT2 subgroups (one trended toward NT1, the other toward IH), and two IH subgroups with differences in hypnodensity, qEEG, and spindle characteristics. Conclusion Our exploratory findings demonstrate strong diagnosis classification performance from nPSG data alone, more easily distinguishing NT2 from CCs than from IH, and IH from CCs. The distinct NT2 subgroups suggest heterogeneity within NT2; further research is warranted to explore these patterns. Statement of Significance Accurate diagnoses of narcolepsy types 2 (NT2) and 1 (NT1) and idiopathic hypersomnia (IH) remain challenging due to overlapping symptoms. We developed a machine learning model using drug-free nocturnal polysomnography data to automatically differentiate NT2 from clinical controls, NT2 from IH, and IH from clinical controls, with high accuracy. Our model leverages a rich set of sleep features, including spindle and quantitative electroencephalogram (qEEG) metrics. Furthermore, our analysis revealed distinct sleep phenotypes within each diagnosis, suggesting subtypes with varying levels of sleep disturbance and differences at qEEG and spindle levels. These findings provide a novel approach to classifying central disorders of hypersomnolence and suggest disease heterogeneity, which could lead to more accurate and timely diagnoses and personalized treatment strategies.
Gong et al. (Tue,) studied this question.