Data-Driven Prediction of Fatigue in Parkinson's Disease Patients.

Numerous non-motor symptoms are associated with Parkinson's disease (PD) including fatigue. The challenge in the clinic is to detect relevant non-motor symptoms while keeping patient-burden of questionnaires low and to take potential subgroups such as sex differences into account. The Fatigue Severity Scale (FSS) effectively detects clinically significant fatigue in PD patients. Machine learning techniques can determine which FSS items best predict clinically significant fatigue yet the choice of technique is crucial as it determines the stability of results. 182 records of PD patients were analyzed with two machine learning algorithms: random forest (RF) and Boruta. RF and Boruta calculated feature importance scores, which measured how much impact an FSS item had in predicting clinically significant fatigue. Items with the highest feature importance scores were the best predictors. Principal components analysis (PCA) grouped highly related FSS items together. RF, Boruta and PCA demonstrated that items 8 ("Fatigue is among my three most disabling symptoms") and 9 ("Fatigue interferes with my work, family or social life") were the most important predictors. Item 5 ("Fatigue causes frequent problems for me") was an important predictor for females, and item 6 ("My fatigue prevents sustained physical functioning") was important for males. Feature importance scores' standard deviations were large for RF (14-66%) but small for Boruta (0-5%). The clinically most informative questions may be how disabling fatigue is compared to other symptoms and interference with work, family and friends. There may be some sex-related differences with frequency of fatigue-related complaints in females and endurance-related complaints in males yielding significant information. Boruta but not RF yielded stable results and might be a better tool to determine the most relevant components of abbreviated questionnaires. Further research in this area would be beneficial in order to replicate these findings with other machine learning algorithms, and using a more representative sample of PD patients.