RT Journal Article SR Electronic T1 Drug Amount Prediction in Parkinson’s Disease using Hybrid Machine Learning Systems and Radiomics Features JF Journal of Nuclear Medicine JO J Nucl Med FD Society of Nuclear Medicine SP 2256 OP 2256 VO 63 IS supplement 2 A1 Salmanpour, Mohammad R A1 Hosseinzadeh, Mahdi A1 Bakhtiari, Mahya A1 Gholami, Amir Reza A1 Ghaemi, Mohammad M A1 Nabizadeh, Amir Hossein A1 Rezaeijo, Seyed Masoud A1 Rahmim, Arman YR 2022 UL http://jnm.snmjournals.org/content/63/supplement_2/2256.abstract AB 2256 Introduction: Parkinson’s disease (PD) is progressive and heterogeneous. Predicting and personalizing drug amounts consistently to treat PD patients holds significant promise to enhance chances of successful temporal symptomatic therapy or significantly diminish PD symptoms and adverse drug reactions. In this study, we aim to predict the amount of levodopa prescribed by physicians and its incremental dose in the future using Hybrid Machine Learning Systems (HMLS). Radiomics features (RF) provide a more accurate analysis of imaging data and are employed in this study in addition to others, including clinical features (CF) and conventional imaging features (CIF). To our knowledge, no study has focused on the prediction of levodopa and its incremental dose in different years.Methods: We selected 264 patients from the Parkinson's Progression Markers Initiative database. We extracted 950 features, including CFs, CIFs, and RFs extracted from SPECT image segmented regions of interest (ROI: both left and right caudate and putamen; delineated via MRI) with RFs extracted using the standardized SERA software. We generated 7 datasets: (i, ii, iii) data in year 0, 1, or 2; (iv, v) longitudinal datasets in years 0&1 and years 0&1&2, and (vi, vii) timeless datasets in years 0&1 and years 0&1&2. Timeless datasets enabled building an effectively larger dataset, as each patient was listed separately multiple times with data from different years. We selected 8 outcomes (O) to predict, namely: O1) patients being on/off drug in year 1, O2) amount of dose in year 1, and O3-8) increase in drug amount in different years such as from 1st to 2nd, 2nd to 3rd, 3rd to 4th, from 4th to 5th, from 1st to 4th, from 1st to 5th year. For the first 3 outcomes (O1-3), we considered four datasets involving years 0 and 1, while other outcomes (O4-8) considered all datasets. All features in each dataset were normalized by z-score techniques. 80% of all datapoints were used for HNLSs to optimize and select the best model based on maximum performance resulting from 5-fold cross-validation. Subsequently, the remaining 20% were used for external testing of the selected model. A range of optimal algorithms was pre-selected amongst various families of learner algorithms. HMLSs included 10 feature extraction (FEA) and 9 feature selection algorithms (FSA) followed by 10 regressors (for prediction of O2-8) and 9 classifiers (for prediction of O1) optimized by 5-fold cross-validation and grid-search).Results: To predict on/off drug status (O1), HMLS, including Random Forest followed by ReliefA applied on the timeless dataset (year 0&1&2) had the highest accuracy ~ 88.5% ± 2.20% and external testing of 95.8. Further, for prediction of drug-amount in year 1 (O2), HMLS including K-Nearest Neighbor Regressor (KNN-R) linked with Minimum Redundancy Maximum Relevance Algorithm (MRMR) applied on the timeless dataset (year 0&1&2) arrived at a mean absolute error (MAE) ~ 47.1 ± 13.6 (outcome range [30.3:850 mgr]) and external testing of 31.9. For prediction of dose increments (O3-8), HMLSs: Unsupervised Feature Selection with Ordinal Locality+KNNR, ReliefA+KNNR, ReliefA+KNNR, Local Learning-based Clustering Feature Selection+KNNR, MRMR+KNNR and MRMR+KNNR applied to timeless datasets resulted in MAEs of 0.42±0.18, 0.10±0.09, 0.04±0.01, 0.24 ± 0.15, 0.25 ± 0.05 and 0.33 ± 0.26 (outcome range [0.23:29.7]), respectively. Moreover, their external testing confirmed our findings. As shown, the performances resulting from the timeless datasets outperformed performances resulting from other datasets. As indicated, KNNR is able to link with various FSAs to enhance performance. Conclusions: We demonstrated that combining clinical features and imaging features as well as employing appropriate HMLSs, significantly improve prediction performances. Furthermore, we showed that using timeless datasets (effectively increasingly data set sizes) enabled outperformance relative to other datasets in predicting outcomes.