Christabelle Pabalan
Portfolio  /  2026

Christabelle
Pabalan

I enjoy building for underserved groups and for causes that impact our quality of life, whether through AI systems or creative pursuits, such as poetry, writing, and dance.

LEAD DATA SCIENTIST  ·  Qventus
Based in Studio City, CA
Email christabellepabalan@gmail.com
LinkedIn christabellecp
GitHub christabellecp
Writing Medium · TDS
Christabelle Pabalan
About

I'm a Lead Data Scientist at Qventus, where I build AI systems for perioperative care coordination.

Before that, I worked at Headspace across clinical NLP and recommender systems, extracting mental health indicators and behavioral signals from therapy chat data, and building personalization systems for content discoverability and provider matching.

My research on predicting Alzheimer's cognitive scores from a single baseline MRI and demographic data was published in Nature Aging (2026). I also hold a publication in recommender systems from my time at Headspace.

My academic background is in cognitive science, machine learning, and neural computation; prior projects include neural signal processing of ECoG data, cancer prognosis from microRNA biomarkers, and network analysis of social influence in media.

Outside of work, I write poetry, run a little data science Etsy shop, dance and love to read about all things health, neuroscience, philosophy, etc.

Selected Work
A non-exhaustive index of things I've built, grouped roughly by the question each one was trying to answer.
Chapter 01 / Mental Health ML at Headspace

Reading chat data & personalizing care for anxiety, depression, and the things in between.

Clinical NLP Mental Health Longitudinal

Goal Progress & Mental Health Indicators from Chat Data

What members and therapists say in chat tells you what's working, and what isn't.

NLP systems that extract structured signal from longitudinal therapy chat data. Turning unstructured language into clinical and behavioral indicators usable by care teams and downstream personalization surfaces.

NLP pipeline extracting anxiety and depression indicators, activity mentions, and goal progress from therapy chat data

Anxiety & Depression Indicators

Identifying symptom signals in therapy and psychiatry chat data, surfaced as structured indicators for clinical context and member-facing surfaces.

Physical Activity Tracking

Extracting activity mentions across chat history to build behavioral trajectories per member.

Goal Progress Across Time

Surfacing goals stated, milestones met, and goal evolution longitudinally across sessions.

Recommender Systems Sequential Rec Personalization

Recommender System & Content Discoverability

Personalization that learns the time intervals between sessions, not just the session counts.

Built and extended the content recommendation system for Headspace's meditation and sleep catalog. Modeled temporal listening patterns while balancing relevance with deliberate exposure to unfamiliar content.

Semantic sampling workflow: TiSASRec predicts the next item, LaBSE embeds it, cosine similarity ranks catalog items, and sampling replaces it with a semantically similar, novel recommendation

TiSASRec Wrapper

Adapted Time-interval Self-Attentive Sequential Recommendation for the meditation and sleep catalog. Modeled not just what members listened to, but when, and how those intervals shaped engagement with future content.

Novelty & Serendipity

Quantified novelty in recommendations to keep members from getting trapped in narrow content loops that flatten the value of meditation tools.

+12% Discoverability
+3% Engagement
Provider Matching Ranking Conversion

Matching Patients to Therapists

Provider matching where the goal isn't a click, it's an appointment that actually happens.

Modeling and ranking work on Headspace's therapy and psychiatry provider matching. Optimized the funnel from initial intake all the way through to delivered appointment, not top-of-funnel engagement.

Provider matching funnel ranking therapists and psychiatrists from initial intake through to delivered appointment
+11% Delivered appts
−11% Unbooked rate
Topic Clustering Content Strategy NLP

Content Gap Discovery via Therapist Link Analysis

What therapists send between sessions tells you exactly what content you should have built.

Analyzed the URLs therapists shared with patients to evaluate how much was sourced from in-house Headspace content versus external sites. Clustered topics that external links covered to surface concrete gaps in the in-house catalog. The analysis circulated widely across the company and directly shaped in-house content production priorities.

Clustering external links therapists shared with patients to identify content gaps in the in-house catalog
Chapter 02 / Clinical AI & LLM Systems

Building production AI for surgical and perioperative care.

LLM Evaluation Framework pipeline: Curate, Generate, Evaluate, Optimize, loop
LLM Eval Healthcare AI Continuous Improvement

Production LLM Evaluation Framework

A loop that lets clinical AI be iterated on safely, and audibly.

Designed and built the production evaluation framework for LLMs used in clinical decision support. Four interlocking layers feed a continuous iteration loop, with precision, recall, and hallucination metrics anchored to clinical ground truth.

Golden Dataset Curation

Coverage-aware sampling across conditions, demographics, and rare cases. The ground truth all other layers measure against.

Scoring & Human Validation

Automated LLM-as-judge scoring against curated rubrics catches regressions early. High-uncertainty cases route to clinical reviewers, where disagreement is treated as the most valuable signal, not noise.

Autonomous Prompt Optimization

Structured search over prompt templates against fixed golden sets, replacing manual prompt tuning.

Hybrid ensemble model architecture: EHR inputs to three surgical outcome models to calibrated risk scores
Hybrid ML Ensemble Perioperative

Hybrid Ensemble Models for Surgical Outcomes & Efficiency

Reading structured EHR data and unstructured clinical notes as one problem.

A line of work applying hybrid ensemble approaches to perioperative risk and surgical workflow efficiency, combining structured EHR signals with unstructured clinical language. Deployed across partner health systems for proactive intervention.

Surgical Cancellation Prediction

Identifying which planned surgeries are at risk for cancellation, and why. Accuracy 0.67 → 0.79, specificity 0.71 → 0.91, precision 0.49 → 0.74.

ASA Physical Status Classification

Patient risk prior to surgery prediction. +152% lift over random, 97% of interrater agreement ceiling. Under revision at JMIR.

Screening Duration Modeling

Predicting and reducing waste in pre-surgical screening workflows. 83% waste reduction, 10.8 hrs saved per 100 patients.

Chapter 03 / Earlier Research & Visualization

Cognitive science meets applied ML.

Brain MRI segmentation predictions in horizontal, coronal, and sagittal views
Multi-task Ensemble Alzheimer's

Multi-task Alzheimer's Cognitive Score Prediction

Borrowing strength across cognitive tasks to predict each one better.

A novel ensemble and multi-task learning approach predicting cognitive functioning scores in Alzheimer's patients, published in Towards Data Science.

Read on Towards Data Science
Group 4 analysis: 13 of 19 symptoms worsened at higher rates than the cohort average during COVID-19
Clustering Parkinson's COVID

Clustering Parkinson's Patients during COVID-19

What a pandemic did to symptom trajectories, and to the cohorts you can build from them.

Unsupervised clustering analysis on Parkinson's patient symptom data during the COVID-19 pandemic, published in Towards Data Science.

Read on Towards Data Science
Brainwave decomposition of temporal lobe signal, awake vs anesthesia
EEG Time-Frequency Consciousness

Time-Frequency Analysis: Oscillatory Basis for Consciousness

Looking for the rhythms a brain holds when it knows it's awake.

Time-frequency decomposition of EEG signals investigating oscillatory neural correlates of conscious experience, undergraduate research in cognitive science.

Read the project
Redesigned Home Health Compare dashboard with bar charts and donut charts
Product Design Home Health UX

Redesigned Home Health Compare

Compliance data, but legible to the people whose care it describes.

A product concept (Haoma Connect) redesigning the CMS Home Health Compare interface to make agency quality data accessible to families choosing home health providers.

Read the project
Publications
Peer-reviewed research spanning clinical AI, neuroimaging, and recommender systems. Published in Nature Publishing Group and indexed journals.
Nature Aging · 2026

Predicting Categorical and Continuous Alzheimer's Disease Outcomes from a Single MRI Scan

D Ma, C Pabalan, A Rajagopal, A Akanksha, Y Interian, Y Yang, A Raj

Nature Aging
Vol. 6, pp. 1121-1137
Open Access
IF: 25.0

Predicts Alzheimer's cognitive scores up to 36 months out from a single baseline MRI scan, no PET, longitudinal data, or cognitive assessments required. State-of-the-art results on both segmentation and cognitive score prediction, for a disease affecting 50M+ people worldwide.

My contribution Led feature design and modeling over several years, including the ensemble architecture combining imaging and tabular modules. The key unlock was treating demographic features as cognitive correlates rather than mere covariates. Also contributed code, model execution, and data preprocessing.
25.0
Impact Factor
83
Altmetric
IJACSA · 2023

Semantic Sampling: Enhancing Recommendation Diversity and User Engagement in the Headspace Meditation App

RS Rajput, C Pabalan, A Chaturvedi, P Kulkarni, AB Headspace

International Journal of Advanced Computer Science & Applications
Vol. 14 (11) · 2023
My contribution Recommender systems research from my time at Headspace, on semantic sampling to improve content diversity and long-term user engagement in the meditation catalog.
RecSys
Domain
Preprint · 2021

Multi-task Learning and Ensemble Approach to Predict Cognitive Scores for Patients with Alzheimer's Disease

D Ma, C Pabalan, A Rajagopal, A Akanksha, Y Interian, Y Yang, A Raj

bioRxiv · 2021.12.08.471856
Precursor to Nature Aging (2026)
2021 → 2026
Published
Writing
Long-form essays on the seams where neuroscience, AI, and the things humans actually feel touch each other, published across Towards Data Science, CognitiveCuriosities, AI Advances, and Towards AI.
Ventures
Things I'm building outside of work, most of them at the intersection of creative practice and the small, slow operations of running something yourself.
Poetry Published 2026

Rescue A Whisper

A debut poetry collection that follows 'The Change Cycle' and explores themes of grief, abuse, faith, and wrestling with guilt.

Available via Amazon, IngramSpark, Barnes and Noble and select indie bookstores in Southern California.

Visit rescueawhisper.com
Candle Brand Hand-poured

Coconut Cottagehouse

Small-batch candles, made with my mom.

A hand-poured candle brand built on clean ingredients (coconut apricot wax, cotton and wooden wicks) and honest scent direction. Ran with my mom; available via Shopify and select retail partners.

Shop the candles
Merch Print-on-demand

SeizeTheData

Data science, but make it wearable.

A data-themed apparel and accessories shop on Etsy / Printify. Originated from a class hoodie contest during my master's program.

Shop on Etsy
Handwritten notes on decision boundaries, recommender systems, supervised learning, and error metrics
Handwritten notes on network diameter, time series, singular value decomposition, and databases
Notes Instagram

DataDiagrams

An ongoing Instagram series of handwritten notes covering ML concepts: decision boundaries, recommender systems, SVD, time series, error metrics, network theory. Originally a study habit; accidentally became a public archive.

Contact

Always up for a good conversation about data ethics, neuroscience, or poetry (or all three!).