Age-adjusted neurofilament light chain cutoffs for diagnosing neurodegenerative dementia: a two-threshold machine learning approach

Clinical implementation of neurofilament light chain (NfL), a biomarker of neurodegeneration, remains challenging due to absence of reliable cutoffs and influence of confounding factors, particularly age. We aimed to develop an age-adjusted, two-threshold classification framework to support clinical interpretation of NfL in the neurodegenerative dementia diagnosis. We retrospectively enrolled subjects with cognitive/behavioral disturbances and cognitively unimpaired controls (CTR). Participants underwent a baseline diagnostic workup, including at least a neuropsychological assessment, blood test, CSF and serum NfL measurement, and MRI, and were followed up for 2 years. At follow-up, they were diagnosed with either a neurodegenerative dementia [Alzheimer's Disease (AD), Frontotemporal Dementia (FTD), or Lewy Body Dementia (LBD)] or a non-neurodegenerative condition [Other Etiology (OE)]. AD, FTD, and LBD were grouped as Neurodegenerative (NDG), while OE and CTR were grouped as Non-Neurodegenerative (non-NDG). Bayesian regression models assessed the effects of age, disease duration, renal function, and gender on NfL. A weighted support vector machine (SVM) with leave-one-out cross-validation defined age-adjusted cutoffs using a two-threshold strategy constraining sensitivity ≥ 85% and limiting the intermediate zone. We included 217 subjects (158 NDG, 59 non-NDG). Serum and CSF NfL levels were significantly higher in NDG. Age strongly increased NfL levels, particularly in serum where diagnostic separation declined beyond age 70. The SVM-based model applied to CSF NfL defined continuous, age-adjusted, two-threshold cutoffs, identifying low-risk, high-risk, and intermediate zones. The proposed framework for CSF NfL provides a clinically oriented decision-support tool to stratify the likelihood of underlying neurodegeneration and guide diagnostic workup in cognitive neurology settings, pending external validation.