Quantification of identifying cognitive impairment using olfactory-stimulated functional near-infrared spectroscopy with machine learning: a post hoc analysis of a diagnostic trial and validation of an external additional trial

Abstract

Background We aimed to quantify the identification of mild cognitive impairment and/or Alzheimer's disease using olfactory-stimulated functional near-infrared spectroscopy using machine learning through a post hoc analysis of a previous diagnostic trial and an external additional trial. Methods We conducted two independent, patient-level, single-group, diagnostic interventional trials (original and additional trials) involving elderly volunteers (aged > 60 years) with suspected declining cognitive function. All volunteers were assessed by measuring the oxygenation difference in the orbitofrontal cortex using an open-label olfactory-stimulated functional near-infrared spectroscopy approach, medical interview, amyloid positron emission tomography, brain magnetic resonance imaging, Mini-Mental State Examination, and Seoul Neuropsychological Screening Battery. Results In total, 97 (original trial) and 36 (additional trial) elderly volunteers with suspected decline in cognitive function met the eligibility criteria. The statistical model reported classification accuracies of 87.3% in patients with mild cognitive impairment and Alzheimer's disease in internal validation (original trial) but 63.9% in external validation (additional trial). The machine learning algorithm achieved 92.5% accuracy with the internal validation data and 82.5% accuracy with the external validation data. For the diagnosis of mild cognitive impairment, machine learning performed better than statistical methods with internal (86.0% versus 85.2%) and external validation data (85.4% versus 68.8%). Interpretation In two independent trials, machine learning models using olfactory-stimulated oxygenation differences in the orbitofrontal cortex were superior in diagnosing mild cognitive impairment and Alzheimer's disease compared to classic statistical models. Our results suggest that the machine learning algorithm is stable across different patient groups and increases generalization and reproducibility.