Education

This project addresses a clinically relevant problem in pediatric gait analysis by leveraging deep learning-based 3D pose estimation. The work is hosted at **SCAI-Lab** in collaboration with the **AIT Lab**, both of which focus on cutting-edge AI methods in healthcare. **Main Objectives:** - Develop a robust pipeline to benchmark 2D-to-3D pose estimation models on a CP dataset. - Fine-tune the top-performing model using data from children with CP. - Analyze joint angle time-series to evaluate the impact of orthosis in comparison with both healthy and non-orthosis CP gait patterns. **Technical Highlights:** - Evaluation of several 2D detection backbones (Detectron2, AlphaPose, MediaPipe, etc.) and 3D lifting models (VideoPose3D, MotionBERT, MotionAGFormer, BlazePose3D). - Application of clustering and deep learning methods (MVTS, LSTM-FCNN, MLP) for multivariate time-series analysis. - Potential for scientific publication depending on the results. **Ideal Candidate:** We are looking for a motivated Master’s student with skills in data processing (especially video/MoCap) and deep learning using Python/PyTorch. Students from backgrounds in Computer Science, Information Technology, Electrical Engineering, and Health Science and Technology are especially encouraged to apply. The project requires a 6-month full-time commitment.

The primary goal of this project is to evaluate the impact of orthotic devices on gait patterns in children with cerebral palsy using 3D pose estimation techniques. To achieve this, the following sub-goals are defined: - **Develop a 3D Pose Estimation Pipeline:** Construct a robust pipeline that combines 2D detection with 3D lifting models suitable for clinical gait analysis in CP. - **Benchmark Pose Estimation Models:** Systematically evaluate combinations of 2D and 3D models (e.g., AlphaPose + VideoPose3D, RTMPose + MotionBERT) to identify the best-performing setup on the CP dataset. - **Fine-Tune for CP Gait Data:** Adapt the top-performing model to CP-specific gait features by fine-tuning it on annotated clinical video data. - **Generate Joint Angle Time-Series:** Convert 3D keypoint predictions into joint angles over time to represent motion in a clinically meaningful format. - **Perform Time-Series Analysis:** Use statistical and machine learning methods (e.g., MVTS, LSTM-FCNN, clustering) to compare gait patterns across three groups: healthy, CP without orthosis, and CP with orthosis. - **Enable Clinical Insights:** Derive interpretable results that can inform clinicians about the efficacy of orthotic treatment from a kinematic perspective. - **Document and Disseminate Findings:** Write a comprehensive research article, and potentially prepare a submission for a relevant scientific conference or journal.

The internship/thesis will be co-supervised by: - **Shreyasvi Natraj**: SCAI-Labs snatraj@ethz.ch - **Dr. Diego Paez-Granados:** diego.paez@hest.ethz.ch To apply, please send: - Your CV - A brief cover letter - Preferred start date and confirmation of 6-month availability

A) Front-end HCI/UI Development 1. Requirements elicitation with clinical and DS stakeholders; translate into personas, tasks, and UX flows. 2. Implement a web app (TypeScript/React) with: - Graph editor (nodes, contemporaneous and lagged edges, constraint editor, confidence/provenance, version history). - Proposal pane for algorithmic edges/lags with “accept/edit/reject” and rationale linking. - Execution panel for simulations/interventions and simple counterfactual what-ifs. - Data binding to small demo datasets (CSV/Parquet) and model preview. 3. Performance & quality: incremental layouts, virtualized lists, keyboard shortcuts, accessibility (WCAG), tests (unit/E2E). 4. Documentation and a lightweight design system (components, patterns, interaction guidelines). B) Clinical Validation 1. Draft study protocol (tasks, hypotheses, metrics) and ethics/IRB materials; coordinate approvals. 2. Prepare case vignettes using retrospective data (e.g., MIMIC-IV/HiRID/eICU*) and simulated scenarios. 3. Run pilot, refine, then conduct the study (think-aloud + summative). 4. Metrics: SUS, task completion time, error rate, NASA-TLX, trust/interpretability Likert scales; qualitative theming. 5. Statistical analysis and reporting; actionable design recommendations.

1. Front-end HCI/UI: Design and implement an interactive graph editor and workflow for constructing, annotating, and executing lag-indexed causal models (with provenance and constraints). 2. Clinical validation: Plan and conduct a usability/validation study with clinicians to assess effectiveness, efficiency, and trust, using realistic ICU/ward scenarios and retrospective data.

Please send your CV and the latest transcript of records from my studies to Yanke Li (yanke.li@hest.ethz.ch)

Causal graphical models (CGM) can be determined via randomized controlled experiments (and this is the golden standard to determine the causal relations and quantify the causal effects), learned purely from data[1](i.e., structural learning or causal discovery) or elicited by domain experts based on their knowledge and experiences. It is also feasible to learn the CGM from data with prior expert knowledge when data become not reliable due to insufficient sample size or measurement errors. Early works in causal discovery utilized constraint-based approaches, such as the PC algorithm and the FCI algorithm. These methods exploit conditional independence relationships among variables to infer causal structures. They can restrict the output graph space through conditional independence assumptions or background knowledge about causal relationships [2]. Score-based approaches, such as the GES (Greedy Equivalence Search) algorithm, aim to find the best causal structure that optimizes a scoring criterion, e.g., the BIC (Bayesian Information Criterion) score. These methods often incorporate expert knowledge through prior specifications or domain-specific constraints on the model parameters [3, 4, 5]. Recent advancements in causal discovery involve hybrid approaches that combine elements of constraint-based and score-based methods. These approaches aim to leverage the strengths of both approaches to improve the accuracy and efficiency of causal structure learning. They may incorporate expert knowledge through the integration of domain-specific constraints and scoring functions. In addition to observational data, causal discovery methods have also been developed to handle experimental data, where interventions or randomized experiments are conducted. These methods utilize the causal effects of interventions to infer causal relationships. Expert knowledge can be incorporated by designing well-controlled experiments or interventions based on domain expertise. Despite the growing amount of work on structural learning with expert knowledge in scientific fields, few of them inspect the uncertainty of expert knowledge (what if the expert goes wrong?). This is highly important since in medical applications, causal discovery with expert knowledge should be taken cautiously and sometimes experts can give false knowledge, leading to serious mistakes. An approach considering the expert uncertainty during CGM construction will be more robust to potential bias induced by individuals. This idea is made possible by recent progress toward evaluating and falsifying a given CGM from observational data without ground truth[6]. Therefore, we hope to develop an iterative causal discovery method with experts in the loop for continual interaction and calibration between experts and data. In the meantime, we are also trying to develop an interactive user interface for medical experts to input their knowledge and edit the learned graph. References [1] Glymour, C., Zhang, K., & Spirtes, P. (2019). Review of causal discovery methods based on graphical models. Frontiers in genetics, 10, 524. [2] Flores, M. J., Nicholson, A. E., Brunskill, A., Korb, K. B., & Mascaro, S. (2011). Incorporating expert knowledge when learning Bayesian network structure: a medical case study. Artificial intelligence in medicine, 53(3), 181-204. [3] O’Donnell, R. T., Nicholson, A. E., Han, B., Korb, K. B., Alam, M. J., & Hope, L. R. (2006). Causal discovery with prior information. In AI 2006: Advances in Artificial Intelligence: 19th Australian Joint Conference on Artificial Intelligence, Hobart, Australia, December 4-8, 2006. Proceedings 19 (pp. 1162-1167). Springer Berlin Heidelberg. [4] Perković, E., Kalisch, M., & Maathuis, M. H. (2017). Interpreting and using CPDAGs with background knowledge. arXiv preprint arXiv:1707.02171. [5] Kleinegesse, S., Lawrence, A. R., & Chockler, H. (2022). Domain Knowledge in A*-Based Causal Discovery. arXiv preprint arXiv:2208.08247. [6] Eulig, E., Mastakouri, A. A., Blöbaum, P., Hardt, M., & Janzing, D. (2023). Toward Falsifying Causal Graphs Using a Permutation-Based Test. arXiv preprint arXiv:2305.09565.

1. Literature Review on Causal Discovery: You will study classic and state-of-the-art causal discovery methods (preferably on real datasets with mixed-type variables) that incorporate expert knowledge, leading to a systematic review of causal discovery methods in different categories (constraint-based vs. score-based) with different assumptions (e.g., latent variables, causal mechanism/function forms, and cycles) for different settings (e.g., static data or time series). 2. Data Exploration and Analysis: You will explore analyze and datasets collected for HAPI (Hospital-Acquired Pressure Injury) and Hospital Stay after PI Surgeries. The data record hundreds of patients into dozens of mixed-type variables including patients’ demographic features, lab testing values, and health conditions observed during their stay. Apart from clinical data, you will also have access to multivariate time series extracted from bio-signals on SCI people under intervention experiments. We aim to learn the causal relations among these variables of interest from the above data with experts’ knowledge. Additionally, we will probe some public datasets of other fields such as heart diseases and earth science with ground-truth graphs. 3. Methodology Development and Deployment: You will help develop and validate an iterative algorithm for causal discovery with the expert in the loop against benchmarks using the simulated and real datasets. There are two potential exploring directions: 1. How will the correct expert knowledge help reduce the search space of graphs to improve the accuracy and efficiency of the algorithm? 2. How to encode prior knowledge or design a post-modification strategy with uncertainty from experts’ input to increase the robustness of the algorithm? Based on the methodology, an interactive web/application interface for doctors to input their knowledge and edit the learned graph is expected. If possible, you will assist in deriving the mathematical guarantee for the iterative algorithm in terms of convergence and stability properties. 4. Presentation and Documentation: You will prepare a high-quality manuscript for publication with a clear and scientific presentation of the results and methodology.

1. Gain unique access and first-hand experience in one of the leading institutions on long-term health management - At the Swiss Paraplegic Center at Nottwil. 2. Learn and apply state-of-the-art research methods in Health Data Science.

1. Strong interest and background in (probabilistic) graphical models and causal discovery. 2. Knowledge of virtual environments (conda / docker). 3. Strong experience with Python (preferred). 4. Structured and reliable working style. 5. Ability to work independently on a challenging topic.

Host: Dr. Diego Paez (SCAI-Lab, ETHZ | SPZ) Supervision: Dr. Diego Paez & Yanke Li (SCAI-Lab, ETHZ | SPZ) Please send your CV and the latest transcript of records from my studies to Yanke Li (yanke.li@hest.ethz.ch)