Multinomial logistic regression algorithm for the classification of patients with common parkinsonian syndromes

Parkinsonian syndromes are common neurodegenerative brain disorders. Due to overlapping clinical presentation, their early differentiatiation is challenging, but very important. We developed and tested a multinomial logistic regression algorithm (MLR) for the classification of 18F-fluorodeoxyglucose positron emission tomography images of healthy subjects (CN) and patients with Parkinson disease (PD), multiple system atrophy (MSA) or progressive supranuclear palsy (PSP).

The MLR is based on already established multivariate spatial covariance technique, known as scaled subprofile model with principal component analysis. In MLR we included three classes of patients instead of a single one and used MLR instead of standard logistic regression. We analysed 20 CN, 20 PD, 20 MSA and 20 PSP subjects in MLR and tested it using leave-one-out approach.

The great majority of the subjects were classified correctly. Identification performance was the best for MSA and PSP with AUC = 0.94, while it was slightly lower (AUC=0.91) for PD and the lowest (AUC=0.87) for CN.

We demonstrated the possibility of using MLR for differential diagnosis of parkinsonisms based on subjects’ metabolic brain images. Its good performance warrants a study on a larger cohort and subsequent adaption in clinical practice.