FGAI4H-I-019-A03 E-meeting, 7-8 May 2020 - PowerPoint Presentation

1. FGAI4H-I-019-A03E-meeting, 7-8 May 2020Source:TG-Psy Topic DriverTitle:Att.3 – Presentation (TG-Psy)Purpose:DiscussionContact:Nicolas LangerUniversity of ZurichSwitzerlandTel: +416353414Email: n.langer@psychologie.uzh.ch Contact:Stefan HaufeCharité - Universitätsmedizin BerlinGermanyTel: +49 30 450 639639Email: stefan.haufe@charite.de Abstract:This PPT contains a presentation of I-019-A01-R01.

2. Prediction of Psychiatric Multimorbidity in a Large Pediatric Sample8th meeting of FG-AI4HGenevaMay 7st-8th 2020Dr. Stefan HaufeCharité Universitätsmedizin Berlin Berlin Center for Advanced Neuroimaging (BCAN)Prof. Nicolas LangerUniversity of ZurichDepartment of PsychologyMethods of Plasticity Researcch

3. Data availability: Sample3Training Data:current release: 1602 subjectsAge 5-21 yearsPopulation: typical developing children and children with psychiatric developmental disorders (~70/ multimorbidities)Test Data (November 8th, 2019):Subsample of training data8th release: approx. 400 subjects / year 3Healthy Brain Network (HBN) sampleUpdate: continuation of data collection (currently ~2000 subjects)

4. Data availability: Sample4Training Data:current release: 1602 subjectsAge 5-21 yearsPopulation: typical developing children and children with psychiatric developmental disorders (~70% multimorbidities) 4Healthy Brain Network (HBN) sampleTest Data (November 8th, 2019):Subsample of training data8th release: approx. 400 subjects / yearMaleFemale

5. Data availability: DataPage 55The information about the DSM-V diagnosis, demographics, cognitive and behavioral data will be made accessible through a .csv file. In the context of the present challenge, only resting state EEG data, demographic information as well as extensive cognitive and behavioral measures will be permitted to derive predictive models. DemographicsAge, genderCognitive Datae.g. WISC Behavioral DataQuestionnaires (SWAN)resting EEGRaw dataPreprocessed dataEEG features e.g. theta-beta ratio, alpha asymmetryPossibly T1-weighted MRI imagesSource reconstructionCortical thicknessNeuroimaging devices vastly differ in their acquisition and operating costs as well as practical applicability. While structural and functional MR imaging is very expensive and can only be performed in specialized centers, EEG systems are low-cost and can be used anywhere, including private medical practices. Overall, the cost-benefit ratio of automated diagnoses relying on the combination of behavioral and electrophysiological (EEG) data may be competitive to the standard practice provided the predictions made by such a system are accurate. This consideration is the starting point for the present proposal, which has the central goal of advancing the diagnosis of psychiatric developmental disorders through automated assessment of behavioral and electrophysiological measures.Prediction of DiagnosisDSM-V consensus diagnosisAnnotation Quality:based on the decision of a clinical team all interviews and materials conducted as basis for the DSM-5 consensus diagnosisconducted by licensed clinicians 5

6. Data availability: DataPage 66The information about the DSM-V diagnosis, demographics, cognitive and behavioral data will be made accessible through a .csv file. In the context of the present challenge, only resting state EEG data, demographic information as well as extensive cognitive and behavioral measures will be permitted to derive predictive models. DemographicsAge, genderCognitive Datae.g. WISC Behavioral DataQuestionnaires (SWAN)resting EEGRaw dataPreprocessed dataEEG features e.g. theta-beta ratio, alpha asymmetryPossibly T1-weighted MRI imagesSource reconstructionCortical thicknessNeuroimaging devices vastly differ in their acquisition and operating costs as well as practical applicability. While structural and functional MR imaging is very expensive and can only be performed in specialized centers, EEG systems are low-cost and can be used anywhere, including private medical practices. Overall, the cost-benefit ratio of automated diagnoses relying on the combination of behavioral and electrophysiological (EEG) data may be competitive to the standard practice provided the predictions made by such a system are accurate. This consideration is the starting point for the present proposal, which has the central goal of advancing the diagnosis of psychiatric developmental disorders through automated assessment of behavioral and electrophysiological measures.Cognitive & Behavioral Data:DemographicsCognition / Intelligence (e.g. WIAT, WISC-V, NIH-Toolbox)Medical history (e.g. addiction family history)Family structure, stress and trauma (negative life events, parenting)Personality traits (Big 5, self-esteem)Coping Strategies (communication skills, interpersonal factors)Physical measures (e.g. bio-electric impedance analysis, BMI, Metabolic rate, heart rate, blood pressure, height, weight, handedness,…)Social status (SES, parents education, family structure)Nr. of features: ~270 (self-/ parent-/ teacher-report) 6

7. Data availability: Data7The information about the DSM-V diagnosis, demographics, cognitive and behavioral data will be made accessible through a .csv file. In the context of the present challenge, only resting state EEG data, demographic information as well as extensive cognitive and behavioral measures will be permitted to derive predictive models. DemographicsAge, genderCognitive Datae.g. WISC Behavioral DataQuestionnaires (SWAN)resting EEGRaw dataPreprocessed dataEEG features e.g. theta-beta ratio, alpha asymmetryPossibly T1-weighted MRI imagesSource reconstructionCortical thicknessNeuroimaging devices vastly differ in their acquisition and operating costs as well as practical applicability. While structural and functional MR imaging is very expensive and can only be performed in specialized centers, EEG systems are low-cost and can be used anywhere, including private medical practices. Overall, the cost-benefit ratio of automated diagnoses relying on the combination of behavioral and electrophysiological (EEG) data may be competitive to the standard practice provided the predictions made by such a system are accurate. This consideration is the starting point for the present proposal, which has the central goal of advancing the diagnosis of psychiatric developmental disorders through automated assessment of behavioral and electrophysiological measures.Raw EEG:5 min. Eyes closed (40 s) & eye open (20 s)128 electrodes (Geodesic EGI system)sampling rate 500 HzNr. of features: ~ 150’000 Relevance of resting state EEG:„We believe these rest activity patterns may reflect neural functions that consolidate the past, stabilize brain ensembles, and prepare us for the future.“ (Buckner & Vincent, 2007)The resting state can be viewed as a starting point from which subsequent cognitions are generated and monitored.“ (Langer et al., 2011)Why EEG:Low costGood psychometric propertiesMobility (wide spread availability)Direct measurement of neuronal activity 7

8. Microstates Analysis in MATLAB(Collaboration: Lars Kai Hansen & Andreas Poulsen)http://www.psychology.uzh.ch/en/chairs/plafor/automagic.htmlAutomagichttps://github.com/methlabUZH/automagicbad-channel reparation, EOG regression, filtering, ICA denoisingpre-processinghdEEGPrerequisite for Reliability: Standardized PreprocessingPrerequisite for Biomarker Research: Reliability of measureshttps://github.com/methlabUZH/automagicAutomagicPedroni, Bahreini Langer, (2018), biorXiv DemographicsAge, genderCognitive Datae.g. WISC Behavioral DataQuestionnaires (SWAN)resting EEGRaw dataPreprocessed dataEEG features e.g. theta-beta ratio, alpha asymmetryPreprocessed EEG:Number of features: ~ 150’000 8Update: all data preprocessed

9. Compute MNI transformation (SET Fiducials)BEM surfacesProject electrodes on scalpCOMPUTE HEAD MODELDeveloping Methods for EEG analysisEEG Connectivity AnalysisHaufe & Langer in prep.EEG Microstates ToolboxPoulsen, Pedroni, Langer, Hansen (2018) 9Update: Almost finished pipeline for functional connectivity features

10. Microstates Analysis in MATLAB(Collaboration: Lars Kai Hansen & Andreas Poulsen)http://www.psychology.uzh.ch/en/chairs/plafor/automagic.htmlAutomagichttps://github.com/methlabUZH/automagicbad-channel reparation, EOG regression, filtering, ICA denoisingpre-processinghdEEGDemographicsAge, genderCognitive Datae.g. WISC Behavioral DataQuestionnaires (SWAN)resting EEGRaw dataPreprocessed dataEEG features e.g. theta-beta ratio, alpha asymmetryEEG featuresFrequency Domain: Frequency Power analysis (e.g. theta/beta ratio; alpha assymetry; 1/f noise, alpha peak) Number of features: ~ 122 Time Domain: Microstates:„MS are stable spatial configurations of the electric field. These spatially stationary microstates might be the basic building blocks of information processing.“ (Lehmann, 1978)Number of features: ~ 40Functional Connectivity: Imaginary part of coherencyTime-reversed Granger causalityNumber of features: ~ 9216 10Update: All features extracted

11. Data Availability 11https://osf.io/ajhgy/wiki/home/Only preprocessed features so farNo raw data

12. Benchmarking1000 Funcional Connectome ProjectBiswal et al., 2010, PNASSmilla Pedroni (2016)Task: prediction of multiple disorders from demographic, phenotypical (cognitive and behavioral) and EEG dataTraining: on public HBN dataBenchmarking: Implementation: participants submit executable codeStandardized input (data folder) and output (binary classification matrix) Container architecture (docker/kubernetes)Free choice of development tools for participantsSafe for organizersCloud computing: GCP/AWS or similarChallenge platform: crowdai.org/Kaggle etc. 12on future releases of HBN data sets (approx. 500 subjects / year)

13. Performance metrics111111000011111110101D disordersN subjectsYtrue : true test labels011001000111010110111D disordersN subjectsYpred : predicted labelsMain metric (used for ranking): multi-task accuracySecondary metrics: F1-score, sensitivity, specificity, precision, recallMulti-task metrics for continuous labels (severity scores) available. 13

14. Timeline1000 Funcional Connectome ProjectBiswal et al., 2010, PNASSmilla Pedroni (2016)Idea: continuous prediction challengeParticipant teams can refine and upload containers any timeBenchmarking of most recent containers each time new data are releasedTime stamp system allows public release of test set without delayTracking progress over time as new releases become available201920202021Team 1Team 2Team 3Initial training phase 14

15. Timeline1000 Funcional Connectome ProjectBiswal et al., 2010, PNASSmilla Pedroni (2016)Idea: continuous prediction challengeParticipant teams can refine and upload containers any timeBenchmarking of most recent containers each time new data are releasedTime stamp system allows public release of test set without delayTracking progress over time as new releases become available201920202021Team 1Team 2Team 3Initial training phaseNew data releaseEligible for benchmarking 15

16. Timeline1000 Funcional Connectome ProjectBiswal et al., 2010, PNASSmilla Pedroni (2016)Idea: continuous prediction challengeParticipant teams can refine and upload containers any timeBenchmarking of most recent containers each time new data are releasedTime stamp system allows public release of test set without delayTracking progress over time as new releases become available201920202021Team 1Team 2Team 3Initial training phaseRefinement phase 16

17. Timeline1000 Funcional Connectome ProjectBiswal et al., 2010, PNASSmilla Pedroni (2016)Idea: continuous prediction challengeParticipant teams can refine and upload containers any timeBenchmarking of most recent containers each time new data are releasedTime stamp system allows public release of test set without delayTracking progress over time as new releases become available201920202021Team 1Team 2Team 3Initial training phaseRefinement phase... 17

18. First Benchmarking ResultsThe multi-class, multi-label problem was decomposed into several binary classification tasks.Support Vector Machines (SVM), Logistic Regression (LGR), Random Forest (RF) Supervised Autoencoder (SAE): An autoecoder-based model 18

19. First Benchmarking ResultsThe multi-class, multi-label problem was decomposed into several binary classification tasks.Support Vector Machines (SVM), Logistic Regression (LGR), Random Forest (RF) Supervised Autoencoder (SAE): An autoecoder-based model 19

20. First Benchmarking ResultsThe multi-class, multi-label problem was decomposed into several binary classification tasks.Support Vector Machines (SVM), Logistic Regression (LGR), Random Forest (RF) Supervised Autoencoder (SAE): An autoecoder-based model 20

21. First Benchmarking ResultsAlternative Approach: Data-Driven Subgroups using k-means clustering 21

22. First Benchmarking ResultsAlternative Approach: Data-Driven Subgroups using k-means clustering 22First Benchmarking ResultsFirst Benchmarking Results

23. Next steps:Set up a Kaggle competitionWork on G-014 TDD documentQuantifying uncertainty (Maurice Weber, ETH) 23Dr. Alpha Tom KodamullilFraunhofer Institute for Algorithms and Scientific Computing (SCAI)

24. Child Mind InstituteMichael MilhamCameron CraddockKenneth SchusterErica HoCity College New York Simon KellyAnnabelle BlangeroIsabel VanegasNatalie SteinemannLucas ParraChildrens Hospital Boston / Harvard Medical SchoolNadine GaabEllen GrantBarbara PeysakhovichJennifer ZukChris BenjaminBryce BeckerNora RaschleChris GorgolewskiAll subjectsUniversity of ZurichAndreas PedroniAmirreza BahreiniLutz JänckeKlaus OberauerJürgen HänggiRoberto Pascual’MarquiMarius TröndleUniversity of GenevaChristoph MichelTechnical University of DenmarkKai Lars Hansen & Andreas PoulsenSwiss Memory ClinicsAndreas MonschMarkus BürgeBirte WeinheimerTHANK YOU FOR YOUR ATTENTION 24Charité Universitätsmedizin BerlinStefan HaufeCe ZhangAffiliations of Prof. Nicolas Langer University of ZurichUFSP Dynamic of Healty AgingZentrum für Neurowissenschaften ZürichDigital Society Initiative (DSI)Center for Reproducible Science (CRS)