Feature engineering-based analysis of DaTSCAN-SPECT imaging-derived features in the detection of SWEDD and Parkinson's disease

Abstract

Scans without evidence of dopamine deficit (SWEDD) refer to patients with a normal dopamine transporter scan who have been clinically considered to have Parkinson's disease (PD). To prevent misdiagnosis between SWEDD and PD, it is imperative to identify the manifestations of SWEDD in the early stages. In this experiment, imaging-based striatum binding ratio (SBR) features of four striatum regions are utilized and obtained from the Parkinson's Progression Marker Initiative (PPMI). After that, by implementing feature engineering, six combinational pairs are created by taking two SBR features at a time using a mathematical-based combination (nCr)technique. Then, by evaluating the maximum and minimum values of features of each pair, the six new features are obtained from each combinational pair based on arithmetic operations. Finally, a total of thirty-six new features have been derived by using these existing four SBR features. These derived 36 features are further applied to various supervised and ensemble classifiers of machine learning for three binary classifications (PD vs SWEDD, PD vs healthy, and healthy vs SWEDD). From the consequences, it has been observed that the Naive Bayes classifier performed well across all classification probabilities and among other classifiers. To differentiate SWEDD from PD, the classifier achieved 98.03 % accuracy, 98.85 % recall, 96.29 % precision, 92.96 % Jaccard_score, 96.29 % F1-score, & 99.34 % AUC. Moreover, the comparison has been made between with and without implementing the feature engineering technique for all three classification probabilities for better interpretation. Thus, it is reasonable to hypothesize that the suggested method, which is based on feature engineering, may not only improve performance but also help medical professionals diagnose diseases at early stages.