WP9: CVD RISK IN OBESE CHILDREN AND ADOLESCENTS MD-Paedigree EU Review Olivier Pauly, Siemens Healthineers
Our vision: a screening approach for identifying obese children with high cardiovascular risk Proposed approach: a sequential strategy for identifying patients at risk through the acquisition of increasingly complex phenotypic measures drawn from very different sources, such as questionnaires, stool and blood samples, clinical assessments and advanced medical imaging.
Our pilot multi-centric study For multi-modal patient modeling To validate the proposed sequential approach, a pilot multi-centric study was conducted to collect a cross-sectional dataset of approximately 160 patients. 2 year follow-ups were performed in approx. 33% of the cases Level I Level II Level III Level IV Level V Questionnaire Microbiome Clinical parameters Imaging Follow-up Genotype
Our pilot multi-centric study For multi-modal patient modeling To validate the proposed sequential approach, a pilot multi-centric study was conducted to collect a cross-sectional dataset of approximately 160 patients. 2 year follow-ups were performed in approx. 33% of the cases Level I Level II Level III Level IV Level V Questionnaire Microbiome Genotype Clinical parameters Imaging Body mass index: BMI Relatives having cardiovascular disease Diet habits Exercise Stress level General health questionnaire Follow-up
Our pilot multi-centric study For multi-modal patient modeling To validate the proposed sequential approach, a pilot multi-centric study was conducted to collect a cross-sectional dataset of approximately 160 patients. 2 year follow-ups were performed in approx. 33% of the cases Level I Level II Level III Level IV Level V Questionnaire Microbiome Populations of intestinal microorganisms at Imaging Follow-up L2, L5 and L6 levels Clinical parameters Genotype SNPs associated to risk factors related to blood pressure, lipids, diabetes2, liver fat,
Our pilot multi-centric study For multi-modal patient modeling To validate the proposed sequential approach, a pilot multi-centric study was conducted to collect a cross-sectional dataset of approximately 160 patients. 2 year follow-ups were performed in approx. 33% of the cases Level I Level II Level III Level IV Level V Questionnaire Microbiome Clinical parameters Fasting glucose Imaging Follow-up Fasting insulin HDL, LDL, triglycerides Systolic and diastolic BP Genotype
Our pilot multi-centric study For multi-modal patient modeling To validate the proposed sequential approach, a pilot multi-centric study was conducted to collect a cross-sectional dataset of approximately 160 patients. 2 year follow-ups were performed in approx. 33% of the cases Level I Level II Level III Level IV Level V Cardiac function parameters extracted from cine MRI Clinical Questionnaire Microbiome Imaging Fat distribution parameters parameters extracted from MRI Dixon and T2* IDEAL Follow-up Genotype
Our pilot multi-centric study For multi-modal patient modeling To validate the proposed sequential approach, a pilot multi-centric study was conducted to collect a cross-sectional dataset of approximately 160 patients. 2 year follow-ups were performed in approx. 33% of the cases Level I Level II Level III Level IV Level V Cardiac function parameters extracted from cine MRI Clinical Questionnaire Microbiome Imaging Fat distribution parameters parameters extracted from MRI Dixon and T2* IDEAL Follow-up Genotype Intermediate outcomes known to be related with frank pathology in later life
Our pilot multi-centric study For multi-modal patient modeling To validate the proposed sequential approach, a pilot multi-centric study was conducted to collect a cross-sectional dataset of approximately 160 patients. 2 year follow-ups were performed in approx. 33% of the cases Level I Level II Level III Level IV Level V Questionnaire Microbiome Clinical parameters Imaging Follow-up 2 years after therapy (diet, balloon, ) Change in BMI Genotype
Extracting intermediate outcomes From imaging data Extract information characterizing cardiac function from cine MRI Extraction of heart anatomy Personalization of whole-body circulation lumped model Cine MRI Strain computation Extraction of motion through polyaffine transform
Extracting intermediate outcomes From imaging data Extract information characterizing cardiac function from cine MRI Cine MRI Extraction of heart anatomy Get access to relevant cardiac function parameters: LV function Arterial stiffening LV wall thickening Strain rate Strain computation Personalization of whole-body circulation lumped model Extraction of motion through polyaffine transform
Extracting intermediate outcomes From imaging data Quantify fat from MRI T2* ideal: liver, subcutaneous, visceral fat
Extracting intermediate outcomes From imaging data Quantify fat from MRI T2* ideal: liver, subcutaneous, visceral fat Get access to relevant fat parameters beyond BMI: Liver fat ratio Subcutaneous fat Visceral fat
Building predictive models Modeling the links between different domains At each stage, predictive models facilitate a risk assessment that determines whether to recall a given patient for the next stage of more advanced examination. Questionnaire Microbiome Clinical parameters Imaging Follow-up Genotype
Building predictive models Modeling the links between different domains At each stage, predictive models facilitate a risk assessment that determines whether to recall a given patient for the next stage of more advanced examination. Questionnaire Microbiome Clinical parameters Imaging Follow-up Genotype
Building predictive models Modeling the links between different domains At each stage, predictive models facilitate a risk assessment that determines whether to recall a given patient for the next stage of more advanced examination. Questionnaire Microbiome Clinical parameters Imaging Follow-up Genotype
Building predictive models Modeling the links between different domains At each stage, predictive models facilitate a risk assessment that determines whether to recall a given patient for the next stage of more advanced examination. Questionnaire Microbiome Clinical parameters Imaging Follow-up Genotype
Building predictive models Modeling the links between different domains At each stage, predictive models facilitate a risk assessment that determines whether to recall a given patient for the next stage of more advanced examination. Questionnaire Microbiome Clinical parameters Imaging Follow-up Genotype
Building predictive models Modeling the links between different domains At each stage, predictive models facilitate a risk assessment that determines whether to recall a given patient for the next stage of more advanced examination. Questionnaire Microbiome Clinical parameters Imaging Follow-up Genotype
Building predictive models Modeling the links between different domains At each stage, predictive models facilitate a risk assessment that determines whether to recall a given patient for the next stage of more advanced examination. Questionnaire Microbiome Clinical parameters Challenges to be addressed: Which type of models? Genotype Imaging How to cope with high-dimensional and heterogeneous data? How to support evidence-based decision making? Follow-up
Our approach: Multi-task deep networks Learning to reduce dimensionality and to predict High risk vs. low risk Decoder Classification Low dimensional representation Encoder
Our approach: Multi-task deep networks Learning to reduce dimensionality and to predict High risk vs. low risk Decoder Classification Low dimensional representation Encoder
Our approach: Multi-task deep networks Learning to reduce dimensionality and to predict High risk vs. low risk Decoder Classification Low dimensional representation Encoder
Our approach: Multi-task deep networks Learning to reduce dimensionality and to predict High risk vs. low risk Decoder Classification Low dimensional representation Encoder
Our approach: Multi-task deep networks Learning to reduce dimensionality and to predict Low dimensional representation Learn patient low dimensional representation using multitask deep neural networks: High risk vs. low risk Learn to reconstruct input Dimensionality reduction objective Classification Learn to predict targets Decoder Supervised classification objective Encoder We ensure the learning of a context-aware and compact representation that permits to infer outcome of interest.
Experimental setup A comparative study Predictive models In our experiments, we compare following models: Deep neural network Deep multi-task autoencoders Logistic regression w. BMI z-score only (BMIz) Random sampling from prior (PRIOR)
True label True label Experimental setup Mesuring performance Metrics We compute the corresponding confusion matrix and following metrics True True Negatives (TN) (TN) False False Positives (FP) (FP) Sensitivity Ratio of nb of true positives over nb of real positives Sens = TP (TP + FN) Specificity Ratio of nb of true negatives over nb of real negatives Spec = TN (TN + FP) False False Negatives True True Positives (FN) (FN) (TP) (TP) Prediction Positive Predictive Value Ratio of nb of true positives over nb of positive detections PPV = TP (TP + FP) F-score Harmonic ( pessimistic ) mean between sensitivity and PPV Fscore = 2.0 Sens PPV (Sens + PPV)
Experimental setup Cross-validation Validation In our experiments, we perform K-folds cross-validation (K=3) Consolidated dataset Normalization Compute metrics Overall Performance metrics Split into K folds Collect folds results Create train, val, test set Train model Evaluate model
Experimental setup Cross-validation Validation In our experiments, we perform K-folds cross-validation (K=3) Consolidated Overall dataset In small-medium Normalization sample size, performance is highly dependent Compute on the choice Performance of the metrics training vs test set. metrics Cross-validation provides an expectation of the performance of the model. Compute p-value Split using into Welch s K t test for comparing distribution Collect of folds performances between 2 models folds results Create train, val, test set Train model Evaluate model
Level I: Questionnaire data For predicting intermediate outcomes INPUT PARAMETERS o dfs_ncore: normalized measure that characterizes the amount of sugar and fat o nssec_mother, nssec_father: social class of the parents, o pss: perceived stress level, o tanner: stage of puberty, o caffeine_mg_d: amount of caffeine (in mg) per day, o ghq: general health questionnaire, permits to assess somatic symptoms, anxiety and insomnia, social dysfunction and severe depression, o wreq: weight-related eating questionnaire o act_overall: level of activity of the patient o BMI-z: body mass index z-score normalized according to the age.
Predicting compliance Defining our target Compliance ability of a hollow organ to distend and increase volume with increasing transmural pressure. Low compliance = stiffness Compromised Windkessel effect that helps converting pulsatile ejection of blood to a more steady flow Increased load on heart, as it needs to perform more work to maintain stroke volume Leads to left ventricular remodelling Increased risk of cardiovascular event Image taken from wikipedia, author Edoarado
Level I: Questionnaire data For predicting aortic compliance
Level I: Questionnaire data For predicting aortic compliance Clear improvement compared to BMI z-score only Significance of results compare to random (P-value of 0,04) Our best model using questionnaire data reaches a sensitivity of 67% and specificity of 80%
Level I: Questionnaire data For predicting aortic compliance Clear improvement compared to BMI z-score only Significance of results compare to random (P-value of 0,04) Our best model using questionnaire data reaches a sensitivity of 67% and specificity of 80% This results suggests that we can identify patients that would require more examination for assessing cardiac function based on cheap questionnaire data.
Level I: Questionnaire data Embedded in 2D Results of embedding 17 dimensions to 2 dimensions
Level I: Questionnaire data Embedded in 2D Results of embedding 17 dimensions to 2 dimensions Higher risk Lower compliance Lower risk Higher compliance
Level II: Microbiome + BMI data For predicting intermediate outcomes INPUT PARAMETERS o L2, L5, L6: families of microorganisms in human gut at different levels o BMI-z: body mass index z-score normalized according to the age.
Predicting fatty liver syndrome Defining our target: liver fat ratio Liver fat ratio: ratio between fat volume and volume of whole organ Fatty liver: reversible disease commonly associated with metabolic syndrome or alcohol. Possible progression towards more severe forms of the disease, i.e. cirrhosis Elevated liver fat ratio we consider a ratio > 5% as being at risk Automated estimation from MRI
Level II: Microbiome + BMI data For predicting liver fat ratio
Level II: Microbiome + BMI data For predicting liver fat ratio Improvement compared to BMI z-score only Significance of results compare to random (P-value of 0,05) Our best model using microbiome + BMI data reaches a sensitivity of 61% and specificity of 88%
Level II: Microbiome + BMI data For predicting liver fat ratio Improvement compared to BMI z-score only Significance of results compare to random (P-value of 0,05) Our best model using microbiome + BMI data reaches a sensitivity of 61% and specificity of 88% This results suggests that we can prevent doing MRI for assessing liver fat based on microbiome and BMI data
Level II: Microbiome + BMI data For predicting liver fat ratio Results of embedding 7 dimensions to 2 dimensions
Level II: Microbiome + BMI data For predicting liver fat ratio Results of embedding 7 dimensions to 2 dimensions Higher risk Higher liver fat ratio Lower risk Lower liver fat ratio
Level III: Microbiome + Clinical data For predicting intermediate outcomes INPUT PARAMETERS o L2, L5, L6: families of microorganisms in human gut at different levels o Clinical data: HOMA-IR, insulin, glucose, HDL, LDL, blood pressure, HR,
Predicting change in BMI at follow up bmi z-score delta vs bmi z-score percentage Delta = bmi-z_follow_up bmi-z_baseline Percent = (bmi-z_baseline bmi-z_follow_up)/bmi-z_baseline Clear response to therapy Weight loss above 0.5 in BMI z-score Weight loss above 10% in BMI z-score
Level III: Microbiome + Clinical data For predicting change in BMI
Level III: Microbiome + Clinical data For predicting change in BMI Improvement compared to BMI z-score only Significance of results compare to random (P-value of 0,004) Our best model using microbiome L5 and clinical data reaches a sensitivity of 67% and specificity of 78%
Level III: Microbiome + Clinical data For predicting change in BMI Improvement compared to BMI z-score only Significance of results compare to random (P-value of 0,004) Our best model using microbiome L5 and clinical data reaches a sensitivity of 67% and specificity of 78% This results suggests that we can infer the outcome of diet based on microbiome and clinical data
Conclusion And outlook Proposed multi-task deep networks are very flexible Their performance is expected to increase with additional data
Conclusion And outlook Proposed multi-task deep networks are very flexible Their performance is expected to increase with additional data Pilot study based on unique deep phenotyping dataset. Results demonstrate feasibility of predicting complex, costly data from simpler, cheaper data This will support applications for future work with larger populations, focused on the most promising variables, needed to build complete sequential screening tools
Thanks a lot For your attention Image courtesy: https://xkcd.com/1838/