AI IN PRECISION HEALTH
Select Computer Science students from Cornell University join select medical students from Weill Cornell Medicine-Qatar for this exclusive opportunity. We're bringing together the best of each campus to develop the future of precision health with artificial intelligence. With support, resources, and travel taken care of you can focus on creating the future.
OCTOBER 10-11, 2025
DOHA, QATAR
24-HOURS OF INNOVATION.
ABOUT
Are you ready to be part of a groundbreaking event that merges the brilliance of computer science from Cornell University with the expertise of medical students from Weill Cornell Medicine-Qatar? Join us for an exciting, global, in-person two-day Precision Health AI Hackathon where innovation meets healthcare, in Doha Qatar!
Why is this happening? Weill Cornell Medicine is committed to advancing healthcare through innovation. This hackathon empowers students to develop AI approaches that address key gaps in patient care and precision health, though interdisciplinary collaboration.
The latest technological advances (AI, Large Language Models, Wearables, Extended Reality) are revolutionizing lifestyles and extending health spans. This unique event, 12 months in the making, brings 12 students from Cornell University-NY to work alongside 18 students from Weill Cornell Medicine-Qatar for one incredible weekend. Interdisciplinary collaboration at its finest. This highly facilitated hackathon includes multiple workshops and tech resources from the worlds finest AI, VR, and medical experts.
AI can advance precision medicine in unprecedented ways. Be part of the vision. Expect the unexpected in this carefully curated opportunity. Unparalleled networking, upskilling, collaboration at its finest. But wait there's more...every participant receives a wearable, with winning team members each receiving a VR set.
WHY PARTICIPATE
THE THEME
AI in Precision Health: The latest technological advances (AI, Large Language Models, Wearables, Extended Reality) are revolutionizing lifestyles and extending health spans. Students utilize the challenges below as a launch point for ideation and inspiration.
1
Federated Learning for Clinical Risk Prediction
Clinical risk prediction models often fail to generalize because they are trained on limited, siloed datasets from a single institution. Access to larger, diverse datasets could improve performance but is constrained by patient privacy and data-sharing regulations. Federated learning (FL) offers a promising solution by allowing models to be trained collaboratively across multiple sites without centralizing sensitive patient data. Leveraging FL in healthcare could unlock better predictive accuracy while preserving privacy, building trust, and enabling broader adoption of AI in clinical practice. Challenge Objective / Question Develop and evaluate a federated learning–based approach for clinical risk prediction. Explore how models trained across distributed datasets can outperform single-site models while preserving privacy. Potential focus area: predicting acute kidney injury (AKI) or other high-risk outcomes in ICU or hospital settings. Dataset & Source •MIMIC-IV (PhysioNet): link – Single-center ICU EHR dataset. •eICU Collaborative Research Database (PhysioNet): link – Multi-center ICU dataset. •HiRID Dataset (PhysioNet): link – Swiss ICU dataset for additional validation. Expected Workflow & Skills •Data preprocessing and harmonization across sites (MIMIC-IV, eICU, HiRID). •Implement federated learning frameworks (e.g., TensorFlow Federated, PySyft, Flower). •Model development: baseline (local) vs centralized vs federated models. •Evaluation of generalizability across sites and patient populations. •Privacy-preserving techniques (e.g., differential privacy, secure aggregation). Deliverables •Federated learning framework with trained predictive models. •Comparative analysis: local vs centralized vs federated approaches. •Generalizability report across multiple datasets. •Reproducible code and documentation. •Short implementation brief on how FL could be integrated into clinical settings. Ethics / Compliance •All datasets are de-identified and open-source; no IRB required. •Explicit documentation of privacy-preserving methods used. •Bias and fairness evaluation across diverse patient subgroups. Evaluation Criteria •Performance: Predictive accuracy, ROC-AUC, sensitivity/specificity across datasets. •Generalizability: Model performance consistency across institutions. •Privacy & Security: Effectiveness of privacy-preserving methods. •Innovation: Novel approaches in FL design or model aggregation. •Reproducibility: Clear documentation, open code, and replicable results. •Clinical Relevance: Applicability of FL in real-world healthcare environments. Expected Impact This challenge highlights how federated learning can bridge data silos in healthcare, enabling more robust, generalizable clinical AI models while maintaining patient privacy. Success could pave the way for collaborative, privacy-preserving AI deployments in hospitals worldwide. MIMIC-IV v3.1 Large database of de-identified health information from patients admitted to Beth Israel Deaconess Medical Center Challenge Provider: Dr. Fei Wang
2
Simulating Clinical Trials with Real-World Data (Target Trial Emulation)
Traditional clinical trials are extremely resource-intensive, costing an average of $2.6 billion and taking years to complete. Meanwhile, vast amounts of real-world data (RWD), such as electronic health records (EHR), contain practice-based insights that can inform treatment effects more efficiently. Target Trial Emulation (TTE) offers a framework to replicate randomized trial principles using observational data. By simulating clinical trials with RWD, researchers can estimate causal effects, validate trial designs, and explore treatment effectiveness in broader, more diverse populations. Challenge Objective / Question Apply the principles of target trial emulation to simulated EHR data to assess the effectiveness of steroid treatment for sepsis patients. Define and execute a trial-like protocol, estimate average treatment effects, and discuss implications for evidence-based medicine. Dataset & Source •Simulated EHR dataset (provided by organizers or derived from public repositories such as MIMIC-IV on PhysioNet). •Includes patient demographics, diagnoses, labs, medications, and outcomes relevant to sepsis and steroid treatment. Expected Workflow & Skills •Understand and frame the clinical question as a target trial protocol. •Define eligibility criteria, treatment strategies, and outcome measures. •Apply causal inference methods (e.g., propensity score matching, inverse probability weighting, doubly robust estimators). •Validate assumptions (e.g., exchangeability, consistency, positivity). •Estimate treatment effects and assess robustness with sensitivity analyses. Deliverables •A defined target trial protocol (eligibility, intervention, outcome, follow-up). •Analysis workflow estimating treatment effects with appropriate causal methods. •Visualization of results (e.g., Kaplan-Meier curves, effect estimates with confidence intervals). •Documentation of assumptions, limitations, and future implications. Ethics / Compliance •Dataset is simulated or de-identified; no IRB required. •Clear explanation of limitations in applying causal inference to observational data. •Transparency around uncertainty and potential biases. Evaluation Criteria •Methodological Rigor: Correct framing of target trial protocol and causal inference approach. •Analytical Accuracy: Robustness of effect estimation and validation of assumptions. •Innovation: Creative use of methods to handle confounding, missingness, or heterogeneity. •Reproducibility: Clarity of code, workflow, and documentation. •Clinical Relevance: Insights generated on steroid effectiveness in sepsis. Expected Impact This challenge demonstrates how real-world data can be leveraged to emulate clinical trials, providing faster, more cost-effective insights into treatment effects. Success will highlight the potential of TTE to complement or accelerate traditional trials, improving evidence generation in critical care and beyond. Challenge Provider: Dr. Fei Wang
3
Predicting Tumor Subtypes in Breast Cancer
Breast cancer treatment decisions often rely on tumor receptor subtypes such as ER, PR, and HER2 status, along with tumor grade and histologic subtype. Today, these are determined invasively via biopsy and pathology, which can delay treatment and increase patient burden. Non-invasive prediction using imaging and clinical data could enable earlier, personalized treatment planning and reduce reliance on invasive procedures. AI models combining MRI scans with clinical metadata have strong potential to advance precision oncology and patient care. Challenge Objective / Question Develop a machine learning model that predicts tumor receptor subtype(s) and/or pathological features using pre-operative breast MRIs and clinical metadata. Participants may: •Use imaging data alone. •Use clinical metadata alone. •Combine both modalities for multi-modal fusion. Dataset & Source •Duke Breast Cancer MRI Dataset (TCIA): link oPre-operative dynamic contrast-enhanced breast MRIs. oClinical variables: age, race, tumor histology, receptor status, etc. Expected Workflow & Skills •Data preprocessing: imaging (normalization, augmentation), metadata cleaning. •Model development: CNNs or transformers for MRI; structured ML models for metadata. •Fusion strategies: integrate imaging + clinical data. •Explainability: visualization tools such as Grad-CAM for MRI interpretability. •Validation: stratified splits, cross-validation, performance benchmarking. Deliverables •Trained predictive model(s) with quantitative performance metrics. •Visual outputs (heatmaps or Grad-CAMs) to explain model decisions. •Comparative analysis of imaging-only vs metadata-only vs multi-modal approaches. •Reproducible code and documentation. •Short proposal for clinical integration and implications. Ethics / Compliance •Dataset is de-identified and publicly available; no IRB required. •Transparency in interpretability to support clinical trust. •Awareness of fairness and potential biases across demographic variables. Evaluation Criteria •Performance: Accuracy, F1, AUC-ROC for classification targets. •Interpretability: Use of visual explanations (e.g., Grad-CAM). •Innovation: Novel approaches to fusion or modeling strategies. •Clinical Relevance: Applicability to personalized treatment planning. •Reproducibility: Clear documentation, code quality, and robustness. Stretch Goals (Optional) •Heatmap localization of tumor regions using weak supervision. •Comparative benchmarking: metadata-only vs imaging-only vs fused models. •Prototype of a simple clinician-facing interface for predictions. Expected Impact A successful solution will demonstrate how AI can combine imaging and clinical data for non-invasive tumor profiling, advancing personalized oncology, and reducing diagnostic burden. Challenge Provider: Dr Ahmad Serag
4
Medical Imaging: Automatic PCOM Classification
Polycystic Ovary Morphology (PCOM) plays a central role in the diagnosis of Polycystic Ovary Syndrome (PCOS). Accurate identification of PCOM through ultrasound is essential for guiding hormonal management, fertility planning, and early interventions to reduce long-term reproductive and metabolic risks. Yet, PCOM assessment remains inconsistent due to variable image quality, scanning protocols, and subjective interpretation. Automating PCOM detection with AI can reduce variability, improve diagnostic reproducibility, and allow earlier, more reliable detection in clinical practice. Challenge Objective / Question Develop AI models to automatically and accurately classify ovarian ultrasound images as PCOM or Non-PCOM, reducing subjectivity and enabling more consistent PCOS-related assessments. Dataset & Source •PCOSGen Ultrasound Dataset (Kaggle): link •Format: 2D ultrasound images (JPEG) extracted from videos. •Annotations: Frame-level labels (PCOM vs Non-PCOM). •Size: 4,668 total images (3,627 PCOM; 1,041 Non-PCOM). oTraining: 3,200 images (80% PCOM; 20% Non-PCOM). oTest: 1,468 images (72% PCOM; 28% Non-PCOM). Expected Workflow & Skills •Preprocessing: Image normalization, augmentation, and imbalance handling. •Model development: Apply CNNs, Vision Transformers (ViTs), or Vision–Language Models (VLMs). •Explainability: Use tools like Grad-CAM or LIME to interpret predictions. •Evaluation: Benchmark against prior approaches and push beyond baseline performance. •Analysis: Document error cases and propose strategies to mitigate misclassification. Deliverables •Robust classification pipeline capable of distinguishing PCOM vs Non-PCOM. •Visual performance demonstration of model predictions and interpretability outputs. •Reproducible code with clear documentation and analysis of imbalance/failure cases. •A short clinical context brief: how the model could improve speed and reliability of PCOS diagnosis. Ethics / Compliance •Dataset is de-identified and publicly available; no IRB required. •Fairness checks to ensure the model performs consistently across varied subgroups. •Clear communication of limitations to avoid over-reliance on automated outputs. Evaluation Criteria •Performance: ROC-AUC, PR-AUC, F1, sensitivity, and specificity (balanced across metrics). •Interpretability: Use of visualization tools to explain decisions. •Clinical Relevance: Feasibility of integration into diagnostic workflows. •Reproducibility: Code quality, clarity, and documentation. Expected Impact A successful solution will advance AI-driven PCOM detection, help standardize diagnosis, reduce subjectivity, and support earlier, more reliable PCOS assessments—ultimately improving women’s health outcomes. Challenge Provider: Dr. Rawan Al Saad
5
Predicting the Transition from Gestational Diabetes Mellitus (GDM) to Type 2 Diabetes Mellitus (T2DM)
Gestational diabetes mellitus (GDM) is one of the strongest predictors of type 2 diabetes mellitus (T2DM). Women with GDM have a tenfold increased risk of T2DM, with up to 50% progressing within five years of delivery. In Qatar, the prevalence of GDM among pregnant women is 31.6%, significantly burdening women’s health and the healthcare system. Early risk stratification is critical for deploying preventive interventions, reducing complications, and lowering healthcare costs. AI-powered prediction models can help identify high-risk women, enabling personalized prevention and follow-up. Challenge Objective / Question Can artificial intelligence models accurately predict which women with a history of GDM will develop T2DM, and which clinical, demographic, or metabolic features are most predictive? Dataset & Source 1.Gestational Diabetes Dataset (Amitha Karkera, Kaggle): link – comprehensive dataset with clinical and metabolic features. 2.Diabetes Dataset (Tushar Poddar, Kaggle): link – simplified dataset suitable for baseline model building. Expected Workflow & Skills •Data preprocessing: handling missing values, feature selection, normalization. •Feature engineering: pregnancy-related factors, postpartum markers, sociodemographic variables. •Model development: supervised ML, ensemble models, and/or deep learning. •Model interpretation: explainability methods (e.g., SHAP) to identify key risk drivers. •Validation: cross-validation, external testing, sensitivity/specificity analysis. •Integration framing: pathways for clinical adoption in women’s health programs. Deliverables •Predictive models with performance metrics (ROC-AUC, F1, sensitivity, specificity). •Risk factor analysis report linking features to progression risk. •Reproducible code and documentation. •Deployment brief describing how the model could support preventive care in practice. Ethics / Compliance •All data are de-identified and open-source; no IRB required. •Transparency in risk communication to avoid stigma. •Considerations for fairness across sociodemographic groups. Evaluation Criteria •Predictive Performance: Balanced and clinically useful accuracy, sensitivity, and specificity. •Interpretability: Clear explanation of risk factors and subgroup analysis. •Clinical Relevance: Feasible pathways for preventive interventions. •Technical Rigor: Data handling, validation robustness, and reproducibility. •Ethical Soundness: Privacy, fairness, and risk communication safeguards. •Presentation: Clarity and persuasiveness of results and recommendations. Challenge Provider: Dr. Alaa Ali Abd-Alrazaq
6
Early Prediction of Postpartum Depression (PPD)
Postpartum depression (PPD) is a common and debilitating condition that can arise within weeks to a year after childbirth. Globally, 10–20% of mothers are affected; estimates in Qatar are higher (≈17.6–22.5%). PPD harms maternal health, infant development, and family well-being. Early risk identification enables timely support and prevention, reducing clinical and societal burden. Challenge Objective / Question Develop and evaluate AI models that predict PPD risk using antepartum and postpartum factors. Demonstrate how predictions can inform proactive screening, referral, or digital support pathways. Dataset & Source •Antepartum & Postpartum Depression Dataset (Figshare): https://figshare.com/articles/dataset/Antepartum_Postapartum_Depression_Files/23100746 Expected Workflow & Skills •Data preparation: cleaning, handling missingness, class imbalance. •Feature engineering: socio-demographics, obstetric history, complications, psychosocial factors. •Model development: classical ML and/or modern methods; calibration and thresholding. •Explainability: SHAP/feature importance to identify actionable risk drivers. •Validation: cross-validation and held-out testing; sensitivity analyses. •Deployment framing: integrate outputs into maternal care workflows (screening cadence, alerts). Deliverables •Trained predictive model(s) with performance metrics (e.g., ROC-AUC, PR-AUC, sensitivity/specificity, F1). •Explainability report highlighting key risk factors and fairness checks across subgroups. •Brief implementation brief: proposed clinical workflow, safeguards, and user experience. •Reproducible code and documentation. Ethics / Compliance •Use only de-identified, open-source data; no IRB required for the provided dataset. •Elevated care for mental-health data: clear limits of use, harm-reduction framing, and referral guidance. •Fairness and bias assessment across relevant subgroups; transparent communication of uncertainty. Evaluation Criteria •Predictive performance: Balanced metrics prioritizing clinically useful sensitivity/specificity. •Interpretability & fairness: Clarity of risk drivers; subgroup analyses and bias mitigation. •Clinical relevance: Feasible pathway for screening and follow-up within maternal care settings. •Technical rigor & reproducibility: Clean code, validation design, and documentation. •Ethical robustness: Privacy, consent posture, and harm-minimizing recommendations. •Communication: Clear, concise presentation of methods, results, and integration plan. Challenge Provider: Dr. Alaa Ali Abd-Alrazaq
7
Leveraging Wearable Device Data & Synthetic Clinical Data/ Notes for Predictive Health Insights
Wearable devices provide continuous streams of physiological and activity data, but by themselves often lack the clinical context needed for deeper predictive insights. Combining these datasets with synthetic clinical data (e.g., demographics, notes, lab markers) can create richer, more representative training sets. This enables exploration of stress resilience, chronic disease risk, and preventive interventions, supporting more personalized and predictive health solutions. Challenge Objective / Question Participants will integrate wearable sensor data with synthetic clinical data to build enriched datasets for predictive modeling. Suggested focus areas: 1.How do demographic and clinical factors influence stress resilience? 2.Can combined wearable and synthetic data be used to predict diabetes status or risk of hypoglycemia/glycemia? Dataset & Source •Wearable Device Dataset (PhysioNet): https://physionet.org/content/wearable-device-dataset/1.0.1/ •Synthetic Clinical Data: Self-generated (e.g., ChatGPT, LLaMA, Gemini, etc.) or open resources such as Synthea: https://synthetichealth.github.io/synthea/ Expected Workflow & Skills •Data cleaning & preprocessing of wearable signals. •Synthetic data generation for demographics, notes, labs. •Feature engineering (stress metrics, activity trends, clinical markers). •Exploratory data analysis (EDA) for insight discovery. •ML pipeline development (classification, prediction, forecasting). •Validation & evaluation of predictive performance. Deliverables •Final dataset (real wearable + synthetic data). •Clear documentation of synthetic data generation and integration process. •Visualizations and insights from EDA. •Predictive model(s) with performance metrics. •Short presentation outlining research question, methods, results, and lessons learned. Ethics / Compliance •All data must remain fully de-identified. •Synthetic data must be clearly labeled and documented. •Transparency in generation methods to avoid hidden biases. Evaluation Criteria •Innovation: Novelty in combining wearable + synthetic data for predictive health. •Technical Quality: Data integration, model soundness, and rigor in analysis. •Usability: Clarity of dataset and methods for reproducibility. •Impact: Relevance to stress resilience or diabetes risk. •Ethics & Transparency: Documentation of synthetic data generation and bias mitigation. •Presentation: Clarity of insights and communication of results. Challenge Provider: Dr. Arfan Ahmed
Solutions are evaulated on this criteria during Saturday's Demo Day
SPECIAL GUESTS
We are thrilled to welcome a diverse group of mentors, judges, and speakers who bring a wealth of experience knowledge to the event. Industry leaders, professors, and experts are here to guide participants, share valuable insights, and evaluate groundbreaking ideas. Get ready to learn from the best and take your skills to the next level.
Alaa Abd Alrazaq PhD
Asst Professor of Research in Population Health Sciences
Weill Cornell Medicine - Qatar
Sulaiman Alshakhs MD
Clinical Associate in Psychiatry
Weill Cornell Medicine - Qatar / HMC
Curtis Cole MD
VP & Chief Global Information Officer
Cornell University
Max Renault PhD
Director, Innovations in Global Precision Health
Weill Cornell Medicine-Qatar
Fei Wang PhD
Associate Dean for AI and Data Science, Frances and John L. Leob Professor of Medical Informatics
Weill Cornell Medicine
Arfan Ahmed PhD
Asst Professor of Research in Population Health Sciences - Center Coordinator, AI Center for Precision Health
Weill Cornell Medicine - Qatar
Thurayya Arayssi MD
Vice Dean for Academic and Curricular Affairs
Weill Cornell Medicine - Qatar
Dr Mohamed Elsherif
Post Doc, Qatar Center for Artificial Intelligence
QCRI - Hamad Bin Khalifa University (HBKU)
Javaid Sheikh MD, MBA
Dean, Professor of Psychiatry, of Healthcare Policy & Research
Weill Cornell Medicine-Qatar
Hadeel Zaghloul Phd
Research Associate in Medicine
Weill Cornell Medicine - Qatar
Mohammad Al Hajjaji MD
Attending Physician, General Pediatric
Sidra Medicine Qatar
Johnny Awwad MD
Executive Chair, Women’s Services | Division Chief, Reproductive Medicine
Sidra Medicine Qatar
Judy Gabeskiria
Sr Web Application Developer
Weill Cornell Medicine
Marty Sullivan MPS '18
Principal AI Platform Engineer
Cornell College of Ag & Life Science
Rawan Al Saad PhD
Research Associate in Population Health Sciences
Weill Cornell Medicine - Qatar
Ayham Boucher
Head of AI Innovations
Cornell University
Anna Halama PhD
Assistant Professor of Research in Physiology & Biophysics
Weill Cornell Medicine - Qatar
Rajat Thomas PhD
Assistant Professor of Research in Physiology and Biophysics
Weill Cornell Medicine - Qatar
THE SCHEDULE
SEPT 27
9am EST US
Team Workshop
90-Minutes
Students join this virtual, session to meet their team & ideate.
OCT 4
9am EST US
Fast Start Workshop
120-Minutes
Students join this virtual, session to work with their team developing, designing, and modifying their idea for the hackathon.
OCT 6
5pm EST US
Tech Workshop
90-Minutes
Technical students join to learn about & access to the github repo's, LLMs, and resources. Led by Ayham Boucher, Head of AI Innovation at Cornell Unviersity
NEW YORK STUDENTS
OCT 8
Depart for Airport
1PM
Detailed information provided to Ithaca students, including transportation to airport, flights, required documents, etc.
TRAVEL
16 hours
+ 8 hour time change
OCT 9
Arrival
5:30PM
FRIDAY - OCTOBER 10
Breakfast & Check-in
8:30AM
Welcome + Kick-off
9:00AM
Mentors Available
11:00AM-4:30PM
Team Updates
11:00AM
Friday Prayer
11:30AM - 12:30PM
Lunch Break
12:30PM
Team Updates
2:30PM
Day Concludes
6:00PM
Mentoring Concludes
4:30PM
VIP Dinner
7:00PM
NEW YORK STUDENTS
OCT 12
Day to Explore Doha
OCT 13
Depart for Airport
4:30am
TRAVEL
14 hours
- 8 hour time change
ARRIVAL
JFK 3:30pm
5pm Bus to Ithaca
RESOURCES
Describe your image
Describe your image
BROUGHT TO YOU BY
ORGANIZING TEAM
Max Renault PhD
Director, Innovation in Global Precision Health
WCM-Q
_PNG.png)
Rebecca Joffrey MBA
IT Director, Strategy & Innovation
Cornell A&S
Rahaf Raid MSc
Research Asst, AI Centre for Precision Health
WCM-Q
Ami Stuart MS
Director, Hackathons
Lecturer, Dyson
Entrepreneurship at Cornell
Nour El Messiri MPH
Project Manager
WCM-Q
Alex De Leon
Hackathon Assistant & Researcher
Entrepreneurship at Cornell
Ayham Boucher
Head of AI Innovations
Cornell University
