Cong Pei1,2, Yurong Sun1,2, Jinlong Zhu1,2, Xinyi Wang1,2, Yujie Zhang1,2, Shuqiang Zhang1,2, Zhijian Yao3,4, Qing Lu1,2. 1. School of Biological Sciences & Medical Engineering, Southeast University, Nanjing, China. 2. Child Development and Learning Science, Key Laboratory of Ministry of Education, China. 3. Department of Psychiatry, Affiliated Brain Hospital of Nanjing Medical University, Nanjing, China. 4. Nanjing Brain Hospital, Medical School of Nanjing University, Nanjing, China.
Abstract
BACKGROUND: In order to reduce unsuccessful treatment trials for depression, neuroimaging and genetic information can be considered as biomarkers. Together with machine-learning methods, prediction models have proved to be valuable for baseline prediction. PURPOSE: To propose an ensemble learning modeling framework that integrates imaging and genetic information for individualized baseline prediction of early-stage treatment response of antidepressants in major depressive disorder (MDD). STUDY TYPE: Prospective. SUBJECTS: In all, 98 inpatients with MDD. FIELD STRENGTH/SEQUENCE: 3.0T MRI and gradient-echo echo-planar imaging sequence. ASSESSMENT: Participants were divided into responders and nonresponders based on reducing rates of HDRS-6 after early-stage treatment of 2 weeks. Fourteen brain regions of interest were selected according to previous studies. An ensemble learning modeling framework was used to integrate imaging data and genetic data. STATISTICAL TESTS: Support vector machine (SVM) with linear kernel was utilized to integrate multimode information and then to construct the prediction model. Leave-one-out cross-validation (LOOCV) was used to evaluate the performance. The position characteristics obtained through SVM-RFE (recursive feature elimination) algorithm and LOOCV was considered to compare each feature's relative importance for the prediction model. RESULTS: Compared with the single-level prediction model, the ensemble learning prediction model showed improvement in prediction performance (accuracy from 0.61 to 0.86 with imaging data and genetic data). Integrated with 14 priori brain regions, the region of interest (ROI) map ensemble learning prediction model can achieve a performance that is analogous with the model with information from whole-brain regions (both with accuracy of 0.81). The integration of genetic features further improved the sensitivity of prediction (sensitivity from 0.78 to 0.87 under the ensemble learning framework). DATA CONCLUSION: Our ensemble learning prediction model demonstrated significant advantages in interpretability and information integration. The findings may provide more assistance for clinical treatment selection in MDD at the individual level. LEVEL OF EVIDENCE: 1 Technical Efficacy: Stage 2 J. Magn. Reson. Imaging 2020;52:161-171.
BACKGROUND: In order to reduce unsuccessful treatment trials for depression, neuroimaging and genetic information can be considered as biomarkers. Together with machine-learning methods, prediction models have proved to be valuable for baseline prediction. PURPOSE: To propose an ensemble learning modeling framework that integrates imaging and genetic information for individualized baseline prediction of early-stage treatment response of antidepressants in major depressive disorder (MDD). STUDY TYPE: Prospective. SUBJECTS: In all, 98 inpatients with MDD. FIELD STRENGTH/SEQUENCE: 3.0T MRI and gradient-echo echo-planar imaging sequence. ASSESSMENT: Participants were divided into responders and nonresponders based on reducing rates of HDRS-6 after early-stage treatment of 2 weeks. Fourteen brain regions of interest were selected according to previous studies. An ensemble learning modeling framework was used to integrate imaging data and genetic data. STATISTICAL TESTS: Support vector machine (SVM) with linear kernel was utilized to integrate multimode information and then to construct the prediction model. Leave-one-out cross-validation (LOOCV) was used to evaluate the performance. The position characteristics obtained through SVM-RFE (recursive feature elimination) algorithm and LOOCV was considered to compare each feature's relative importance for the prediction model. RESULTS: Compared with the single-level prediction model, the ensemble learning prediction model showed improvement in prediction performance (accuracy from 0.61 to 0.86 with imaging data and genetic data). Integrated with 14 priori brain regions, the region of interest (ROI) map ensemble learning prediction model can achieve a performance that is analogous with the model with information from whole-brain regions (both with accuracy of 0.81). The integration of genetic features further improved the sensitivity of prediction (sensitivity from 0.78 to 0.87 under the ensemble learning framework). DATA CONCLUSION: Our ensemble learning prediction model demonstrated significant advantages in interpretability and information integration. The findings may provide more assistance for clinical treatment selection in MDD at the individual level. LEVEL OF EVIDENCE: 1 Technical Efficacy: Stage 2 J. Magn. Reson. Imaging 2020;52:161-171.
Authors: J Attal; B Cabarrou; T Valentin; J P Nesseler; E Stoeckle; A Ducassou; T Filleron; S Le Guellec; B Boulet; G Vogin; G Ferron; E Cohen-Jonathan Moyal; M Delannes; C Chevreau Journal: Strahlenther Onkol Date: 2021-10-21 Impact factor: 3.621