◆ RESL 1500 · Undergraduate Research Certificate

Research in Human–Computer Interaction and Clinical Informatics

Deepansh Sharma · Bachelor of Computing Science, 4th Year · Thompson Rivers University

This ePortfolio documents my undergraduate research journey at TRU — from gesture-controlled robotics using Google MediaPipe to clinical data analysis with the MIMIC-III critical care database.

Research Ambassador, TRUUREAP — MIMIC-IIICCECE 2026 (May)
Co-op · Moving Mountains LogisticsProject AURA · MediaPipe Robotics
// 01

Introduction

I am a fourth-year Bachelor of Science student in Computer Science at Thompson Rivers University (TRU), and I have spent the past two years actively engaged in undergraduate research across two distinct domains: clinical data informatics and human-computer interaction. These experiences — combined with my role as a Research Ambassador within TRU’s Undergraduate Research department — have given me a comprehensive understanding of the research process, from formulating a question to communicating findings to a broader audience, and they form the basis of my application for the Undergraduate Research Certificate.

My primary research project was conducted through the Undergraduate Research and Experience Award Program (UREAP), in which I investigated bias in healthcare machine learning models using the MIMIC-III (Medical Information Mart for Intensive Care III) critical care database — one of the most widely used de-identified clinical datasets in health informatics research. Working under the supervision of Dr. Anthony and Dr. Nisha, I designed and carried out a study titled “Developing a Machine Learning Model to Evaluate and Mitigate Bias in Healthcare AI Datasets” that involved processing over 49,000 patient records, applying XGBoost and ensemble methods with demographic reweighting, and interpreting results in the context of existing clinical literature. This project required me to engage seriously with the ethics and practicalities of working with sensitive medical data, and it pushed me to develop research skills I had not previously applied in a real-world setting.

My second major research contribution is a paper I authored on a gesture-controlled robotic arm system built using Google MediaPipe — a framework for real-time hand pose estimation. The project involved designing and building a working prototype that translates hand gestures detected via a webcam into servo motor commands, enabling touchless control of a physical robot arm. This paper was accepted for presentation at the Canadian Conference on Electrical and Computer Engineering (CCECE) 2026, taking place in May 2026 — representing my first peer-reviewed conference publication and public presentation of original research.

In addition to these projects, I serve as a Research Ambassador in TRU’s Undergraduate Research department, where I support the promotion and communication of research opportunities to undergraduate students across campus. This role has deepened my understanding of what undergraduate research looks like across disciplines and has reinforced the value of making research accessible and clearly communicated. Taken together, these experiences demonstrate my sustained engagement with research across its full cycle — from design to dissemination — and qualify me for the Undergraduate Research Certificate.

49,694
Patient Records
0.97
Baseline AUROC
62
Features
CCECE
2026 Conference
// 02

The Standards

Evidence demonstrating engagement across five research competency areas. Each standard has separate evidence even where drawn from the same project.

01 Understanding the Research Process
Evidence

UREAP Project — MIMIC-III Bias Analysis

Through the UREAP program, I independently designed a research study investigating bias in healthcare AI. Under the supervision of Dr. Anthony and Dr. Nisha, I formulated research questions, conducted a literature review, selected appropriate datasets (MIMIC-III), designed a methodology combining XGBoost classifiers with demographic reweighting, and structured a 12-week project timeline. This process required me to move beyond course-level problem sets and engage with the full arc of research — from identifying a gap in existing knowledge to planning how I would address it.

UREAP Project Proposal

My self-authored UREAP proposal defined three research questions: (1) What types of biases exist in MIMIC-III relating to race, gender, and socioeconomic status? (2) How do identified biases impact ML model performance? (3) What mitigation strategies can improve fairness? The proposal included a 12-week timeline, a detailed methodology using counterfactual analysis and intersectional frameworks, and a budget.

📎 UREAP Research Proposal (PDF)

Research Ambassador, TRU Undergraduate Research

In this role, I help undergraduate students across faculties understand what research looks like in their discipline, navigate funding applications (UREAP, NSERC USRA), and connect with faculty mentors. Explaining the research process to students from nursing, business, and arts forced me to articulate the principles of inquiry in discipline-agnostic terms — strengthening my own understanding of how research is structured, communicated, and supported at an institutional level.

02 Evaluating Existing Research
Evidence

Literature Review — Bias in Healthcare AI

As part of my UREAP project, I conducted a self-authored literature review engaging critically with existing research on bias in healthcare ML. I evaluated four peer-reviewed sources: Minot et al. (2021) on gender bias in EHRs, Robinson et al. (2021) on bias in medical language models, Sivarajkumar et al. (2023) on the Fair Patient Model, and Martins et al. (2024) on racial bias in pulse oximeters. The review identified a critical gap: most studies rely on theoretical models without testing mitigation on real healthcare datasets — the gap my research aimed to fill.

Related Work — MediaPipe Robotic Arm Paper (CCECE 2026)

The related work section of my CCECE 2026 paper surveys existing research in real-time hand gesture recognition, human-robot interaction, and pose-estimation frameworks including MediaPipe, OpenPose, and earlier CNN-based classifiers. This section critically positions my approach relative to prior work and identifies the technical gap my prototype addresses.

03 Applying Research Methods
Evidence

Methods — MIMIC-III Clinical Informatics Analysis

This study employed a secondary data analysis methodology using the MIMIC-III v1.4 critical care database. Patient cohorts were constructed by merging PATIENTS, ADMISSIONS, and ICUSTAYS tables, filtering for adults (age 18+), and retaining first ICU stays — yielding 49,694 records with 62 engineered features after one-hot encoding. Data was extracted using BigQuery and Python. A baseline XGBoost classifier was trained, followed by a fairness-aware version using the Kamiran-Calders demographic reweighting method across race-age-outcome groups. Multiple models (Gradient Boosting, Random Forest, AdaBoost, MLP) were evaluated using AUROC, AUPRC, Equal Opportunity Difference, and False Negative Rate disparity. Gender bias was assessed by training separate models on male and female cohorts processed in 500K-record chunks due to server memory constraints.

Methods — Project AURA: Vision-Controlled Robotic Arm

Project AURA employed a distributed three-tier architecture: a React/TypeScript frontend with MediaPipe for real-time hand tracking (21 landmarks at 30 FPS), a Node.js middleware server for angle mapping and OpenAI GPT-4o-mini voice command processing, and an Arduino Uno R4 WiFi controlling 7 MG90S metal-gear servo motors (5 fingers + wrist pitch/yaw). A vector-geometric actuation approach was used instead of inverse kinematics, computing finger flexion angles directly from landmark coordinates. The system was validated through latency benchmarking, 1-hour autonomous loop stability tests, emergency stop response measurements, and user testing across non-technical, engineering, and management profiles.

04 Analyzing and Drawing Conclusions
Evidence

Results and Discussion — MIMIC-III Bias Analysis

The baseline XGBoost model achieved an AUROC of 0.9737 and AUPRC of 0.7959. The fairness-aware model (with demographic reweighting) achieved AUROC 0.9540 and AUPRC 0.7137 — a modest accuracy trade-off for improved equity. Gradient Boosting was the best-performing general model (Test AUC: 0.7358, Accuracy: 88.84%).

The most critical finding concerned False Negative Rates by race: Hispanic/Latino patients showed a 100% FNR — the model missed every single death in this group. Asian patients: 50%, White patients: 31.6%, Black/African American patients: 20%. The Equal Opportunity Difference was 0.800.

Gender analysis showed female patients had higher mortality rates (13.49% vs. 11.36% for males) and slightly lower model AUROC. Insurance analysis revealed Self Pay (20.7%) and Medicare (20.5%) patients had mortality rates over 3x higher than Private insurance patients (6.5%).

Conclusion: Demographic reweighting reduces overall group imbalance but does not eliminate life-critical disparities. Future work must explore adversarial de-biasing and domain-informed model adjustments.

📎 Final Research Report (PDF)
Racial GroupFNRVisual
Hispanic/Latino100%
Unable to Obtain100%
Other66.7%
Asian50.0%
White31.6%
Unknown22.2%
Black/African Am.20.0%

Results and Discussion — Project AURA (CCECE 2026)

Project AURA — a vision-controlled, IoT-enabled, 3D-printed robotic arm — was evaluated across latency, accuracy, and autonomous operation. Key results:

MetricResult
End-to-End Latency171–362ms (sub-500ms)
Gesture/Voice Command Accuracy95% (76/80 commands)
Autonomous Loop Stability (1 hr)99.86% (1,438/1,440 cycles)
Emergency Stop Response185ms mean, 231ms max
Servo Drift (1 hr continuous)< ±2.3°
Payload CapacityUp to 500g
Cloud Data Retention100% fidelity (zero corruption)
Non-Technical User Deploy Time2–3 min (95% success)

The system integrates Google MediaPipe (21 hand landmarks at 30 FPS), OpenAI GPT-4o-mini for voice commands, and Firebase IoT for remote deployment. The “Record, Loop and Deploy” paradigm allows non-programmers to teach robots tasks in seconds via webcam, eliminating proprietary programming languages. Accepted for presentation at the Canadian Conference on Electrical and Computer Engineering (CCECE) 2026.

05 Engaging in Knowledge Mobilization
Evidence

CCECE 2026 — Conference Presentation (May 2026)

I am co-presenting a peer-reviewed paper titled “Project AURA: Vision-Controlled IoT-Enabled 3D-Printed Robotic Arm” at the Canadian Conference on Electrical and Computer Engineering (CCECE) 2026. Co-authored with Gursahib Singh, Yassh Singh, Noori Arora, Vansh Sethi, and supervised by Dr. Anthony Aighobahi.

Authored Paper — Gesture-Controlled Robotic Arm Using Google MediaPipe

Writing this paper required translating a working prototype into a structured academic contribution — situating our design within existing literature, formalizing our methodology, presenting quantitative results, and articulating limitations honestly. The paper underwent peer review and was accepted, meaning our claims and methods were evaluated by external experts in the field before publication.

📎 CCECE 2026 Paper — Project AURA (PDF)

WORKING PROTOTYPE DEMO

Research Ambassador, TRU Undergraduate Research Department

Beyond formal publication, I communicate research to non-specialist audiences through my Research Ambassador role. I run information sessions for undergraduates, explain how to navigate UREAP and NSERC applications, and help students see research as accessible rather than exclusive. This work has taught me that making research understandable to people outside your field is a skill in itself — and one that matters as much as the research itself.

// 03

Reflective Essay

“I was too busy debugging to wonder why the bug existed in the first place.”

Before Research

For the first three years of my degree, I kept research at arm’s length. I saw myself as a builder — someone who writes code, ships features, and moves on to the next problem. Research felt like something that belonged to a different kind of student: the ones who read papers for fun, who lingered after lectures to ask theoretical questions. I assumed building and investigating were separate activities, and I had chosen my side.

The first crack in that assumption came through my role as a Research Ambassador at TRU. Helping students across faculties understand research opportunities forced me to articulate what research actually is — and I realized I could not. When a nursing student asked me how research in computer science works, I stumbled. That embarrassment stayed with me. I started sitting in on presentations, reading abstracts, paying attention. Slowly, I began to see research not as an alternative to building, but as the reason you build anything worth building.

The Turning Point

The UREAP project made this concrete. When I first accessed the MIMIC-III database — 49,000 ICU patient records, each representing someone who was critically ill — I treated it like a Kaggle competition. Extract features, train a model, optimize accuracy, done. My baseline XGBoost hit an AUROC of 0.97 and I thought I was finished.

But when Dr. Anthony asked me to break the results down by race, the picture fell apart.

The model achieved a 100% false negative rate for Hispanic and Latino patients — it missed every single death in that group. A model that looks excellent in aggregate can be catastrophically unfair underneath. That was the moment research stopped being an assignment and became something I needed to understand.

I spent the following months reproducing published studies on racial and gender bias in mortality prediction, running experiments on a remote Linux server that kept killing my processes when I tried to load too much data into memory. I learned to work in 500K-record chunks, to structure reproducible pipelines, to implement fairness metrics I had only read about in papers. The technical skills mattered, but the bigger shift was internal. I stopped asking “what is the accuracy?” and started asking “who does this model fail?”

Why It Matters

Project AURA taught me the other half of the lesson. Where the MIMIC-III work was about uncovering problems in existing systems, the robotic arm project was about creating something new and defending it to strangers. Writing a peer-reviewed paper for CCECE 2026 — co-authored with Gursahib, Yassh, Noori, Vansh, and supervised by Dr. Anthony — required a completely different discipline. Every claim needed evidence. Every design choice needed justification.

Project AURA — Results We Defended

End-to-End Latencysub-500ms
Gesture Recognition Accuracy95%
1-Hour Autonomous Stability99.86%
Non-Technical Deploy Time2–3 min

Writing those numbers into a paper that would be evaluated by experts was the hardest and most rewarding writing I have done.

These experiences have bled into the rest of my life in ways I did not expect. At my co-op with Moving Mountains Logistics, I catch myself applying research thinking to operational problems — asking for data before proposing solutions, questioning assumptions that seem obvious, considering edge cases that a “good enough” approach would ignore. I am not a different person than I was two years ago, but I ask better questions now, and I am more honest about the limits of my answers.

Looking Forward

I plan to pursue graduate study in applied machine learning or health informatics. The MIMIC-III project showed me that there is unfinished, genuinely important work in making AI systems fair for the patients who depend on them. Project AURA showed me that I can take an idea from prototype to publication. And the Research Ambassador role showed me that communicating research clearly is not a soft skill — it is the mechanism by which research becomes useful. I came into this certificate thinking I would document what I had already done. Instead, it clarified what I want to do next: build things that matter, investigate whether they actually work for everyone, and share what I find so others can do the same.

DEEPANSH SHARMA

Bachelor of Computing Science · Thompson Rivers University · Kamloops, BC

linkedin.com/in/deepanshsharmaa  ·  deepanshsharma.trubox.ca

TRU is located on the traditional lands of the Tk’emlúps te Secwépemc within Secwepemcúlucw, the traditional and unceded territory of the Secwépemc peoples.