DE LA RECHERCHE À L’INDUSTRIE Commissariat à l’énergie atomique et aux énergies alternatives www.cea.fr Machine Learning using Neuroimaging for Psychiatric Disorders Edouard Duchesnay, Ph.D., HDR Research director at CEA, NeuroSpin – Université Paris-Saclay, France 

2 1. INTRODUCTION Artificial intelligence + Neuroimaging for mental disorders 2. APPLICATIONS Prognosis of clinical evolution or response to treatment Outline 

3 INTRODUCTION 1. ARTIFICIAL INTELLIGENCE 2. NEUROIMAGING 3. PSYCHIATRY 

4 Artificial intelligence (AI) and Machine Learning (ML) Discover general patterns associated to some semantic from set of observations Input space: similarities are not related to any semantic (cat vs dog) Representation space capture similarities related to some semantic (cat vs dog) dogs cats Now imagine that: – Cats = brain MRI of responder to treatment – Dogs = brain MRI of non-responder to treatment Artificial intelligence? Artificial intelligence → Statistical learning or Machine learning 

5 Psychiatric disorders (simplified vision) [Adam Nature 2013, Craddock, Br. J. Psychiatry 2010] [Solmi Molecular Psychiatry 2021] Bipolar Disorder Depression Schizophrenia Autism spectrum disorders Mood disorders 20.5 20.5 Neurodevelopmental disorders 5.5 Syndromes Symptoms Social cognition emotional response Delusion Disordered thoughts Mood swings Age of onset Causes Genetic Environment (stress) Genetic 

7 Neuroimaging: Brain maturation and mental disorders Synaptogenesis Myelination Synaptic pruning Fetal Childhood (0-12 years) Adolescence (12-20 years) Adulthood (20+) Brain Maturation Mental disorders Genetic vulnerability Abuse Stress Toxics Environment 

8 Brain biomarkers for clinicians Psychiatry lacks objective quantitative measures (such as blood dosage) to guide clinicians in choosing a therapeutic strategy I feel like everyone is staring at me Psychosis ? Bipolar ? Chronic disorder trajectory (risk of, with 80% of chance) Resilience trajectory (risk of, with 20% of chance) 

9 Machine learning for precision psychiatry (personalized medicine) Datasets Heterogeneous patients Deep phenotyping Personalized therapies Stratify patients into subgroups: ● at-risk vs resilient ● responder vs non-responder Biomarqueurs AI models Identify biomarkers of homogeneous subgroups 

10 Bipolar Disorder Depression Schizophrenia Autism spectrum disorders 60-65% Brain imaging encode subtitles markers of psychiatric disorders AI Prediction accuracy Disorder Present: Machine learning can prediction of clinical status from neuroimaging 65-70% [Duchesnay 2012 Neuroimage] [Laidi 2022 Bio. Psy.] 75-80% [De Pierrefeu 2018 Act. Psy. Scand.] [De Pierrefeu 2018 HBM] [Iftimovici 2022 NeuBioRev] 65-75% [Dufumier 2021 MICCAI] 

11 Actual situation: cross-sectional (case/control) datasets Wide phenotyping: Large (1000) longitudinal datasets with individual trajectories vs years 14 16 18 20 Chronic disorder trajectory Resilience trajectory Deep phenotyping: Small (N<100) with PET, advanced MRI, Future: More wide (N>1000) and deep phenotyped longitudinal datasets 

12 Data availability is the primary lock to develop personalized medicine models Collecting large longitudinal datasets is incredibly difficult ● Cost: >20k€ / participants (7 M€ for 170 subjects, ie. >200M€ for 10 000 participants) ● Human resources: recruitment require nurses, physicians, etc. Health care systems cannot implement large scale study Machine learning requires data Learning 

13 Infrastructure to collect more data Prognostic biomarkers of response to lithium in Bipolar disorder. 2018, N~300, F Bellivier Prognostic biomarkers of transition to psychosis patients with first episode psychosis 2020, N~500, MO Krebs French Minds (PEPR, 35M€) Expected in 2022, N~3000, M Leboyer Prognostic biomarkers of clinical evolution in transdiagnostic cohorts of patients with psychiatric disorder NeuroSpin leads data analysis (AI) and management CATI (J. Mangin) for harmonized MRI collection IHU-ICE G Dehaene, T Bourgeron, R Delorme 

14 APPLICATIONS: 1. Prognostic of clinical evolution Asynchronous neural maturation predicts transition to psychosis 2. Prediction of response to treatment Prognosis of response to Lithium in Bipolar Disorder 

15 Clinical objective 1/2: early prediction of detrimental outcome Example: Prediction of transition chronic psychosis Fetal Childhood (0-12 years) Adolescence (12-20 years) Adulthood (20+) Chronic psychosis Toxics Care Treatments 2) Recommend Specific therapeutic strategy 1) AI models to predict future evolution: Identify future subjects at risk Prognostic biomarkers of transition to psychosis patients with first episode psychosis (2020, N~500, PI MO Krebs) AI Resilient First episode psychosis 

16 Asynchronous neural maturation predicts transition to psychosis Anatomical MRI of 82 participants with Ultra High Risk of Psychosis (UHR) (~ 20 years) Transition to psychosis - 27 convertors - 55 Non-convertors One year AI Prediction performances ROC-AUC=80±4% (p < 0.001) BACC=69±5% (p < 0.001) 

17 Asynchronous neural maturation predicts transition to psychosis 1/2. Identify predictive signature ventromedial PreFrontal Cortex right OrbitoFrontal Cortex left Precentral Gyrus Anatomical MRI of 82 participants with Ultra High Risk of Psychosis (UHR) (~ 20 years) Transition to psychosis - 27 convertors - 55 Non-convertors One year AI Prediction performances ROC-AUC=80±4% (p < 0.001) BACC=69±5% (p < 0.001) Predictive signature of psychosis Model Elastic-Net-Total-Variation (Enet-TV) Regions of importance: 1. ventromedial prefrontal cortex (vmPFC): GM ↑ in convertors 2. left precentral gyrus (lPG): GM ↓ in convertors 3. right orbitofrontal cortex (rOFC): GM ↓ in convertors 4. left posterior cingulate gyrus (lpCG), 5. right putamen (rP). 

18 Train brain age predictor using > 2000 anatomical MRI of controls (train: 1605, test: 419). AI Compute brain age gap of of the regions most predictive of psychosis, i.e., evaluate brain maturation vs the general population ventromedial PreFrontal Cortex right OrbitoFrontal Cortex left Precentral Gyrus Asynchronous neural maturation predicts transition to psychosis 2/2. Interpretation: regional maturation analysis 

19 Accelerated brain maturation in converters compared to non-converters, and controls left Precentral Gyrus right OrbitoFrontal Cortex ventromedial PreFrontal Cortex Delayed brain maturation in converters compared to non-converters Asynchronous neural maturation predicts transition to psychosis 2/2. Interpretation: regional maturation analysis 

20 Clinical objective 2/2: Prediction of response to treatment Example: Lithium response of patients with Bipolar Disorder Prognostic biomarkers of response to lithium in Bipolar disorder. (2018, N~300, PI F Bellivier) 1) Scan patient with Bipolar Disorder 2) Lithium treatment 4) AI models to predict response AI 3) Response Yes No 5) Recommend Lithium for potential responders 

21 7Litium brain concentration in patients with Bipolar Disorder Initiate Lithium treatment in 25 patient with BD Measure lithium distribution in the brain → Li concentration in the Hippocampus Future step is Li Concentration map predictive of response to Lithium? Fawzi Boumezber, Frank Bellivier [Stout, Boumezber et al. Biol Psychiatry 2020] 

22 …. Use for clinicians Chronic disorder trajectory (risk of, with 80% of chance) Resilience trajectory (risk of, with 20% of chance) Psychiatry would become closer to other medical specialties Biological-based measurements can help to define a therapeutic strategy Can the introduction of biology-based measures improve the relationship with the patient? Brain signature 

23 Thank you! B Dufumier PhD R Louiset PhD A Iftimovici PH Psy S Petiton PhD B Dollé Engineer P Gori researcher A Grigis researcher F Boumezber researcher J Houenou PUPH Psy P Favre researcher JF Mangin Lab head MO Krebs PUPH Psy M Leboyer PUPH Psy F Bellivier PUPH Psy R Jardri PUPH Psy C Laidi PH Psy M Chupin Engineer P Auriau PhD L Dorval Engineer A Cachia PU Psy A De Pierrefeu, F Hadj Selem, T Lofstead, J Victor, V Frouin, C Philippe, P Ciuciu, D Papadopoulos Close collaborators in psychiatry