Phenotyping of General Surgery Patients for Intraoperative Risk Using Preoperative Waveform Data

Presentation Type: Podium Abstract
Presentation Time: 02:00 PM - 02:12 PM

Primary Track: Data Science/Artificial Intelligence


Preoperative physiological waveforms contain rich biomarker information that can support early diagnosis and more precise surgical risk stratification. Photoplethysmography (PPG) signals, in particular, provide continuous, noninvasive measures of vascular tone and perfusion and have been shown to detect vascular compliance changes and hemodynamic instability. In this study, we applied a biosignal-based foundational model to perform contextual phenotyping of pre-operative PPG signals for patients undergoing general surgery, with the goal of identifying physiologic phenotypes associated with intraoperative and postoperative risk.

We retrospectively analyzed preoperative PPG waveforms from 2,016 general surgery cases in the VitalDB high-fidelity physiologic database. Signals were bandpass-filtered and z-score normalized before generating latent embeddings using an open PPG foundation model fine-tuned on more than 57,000 hours of publicly available PPG data across diverse clinical and wearable datasets. Resulting 512-dimensional embeddings were reduced using UMAP and clustered using silhouette-guided k-means. Cluster phenotypes were evaluated using intraoperative clinical and physiologic metadata. Continuous variables were compared using Kruskal–Wallis tests and categorical variables using Chi-square tests with Bonferroni correction.

Four physiologic phenotypes emerged, each associated with distinct intraoperative profiles. Cluster 2 demonstrated the greatest intraoperative burden, including a doubling of median blood loss (150 mL to 300 mL), higher crystalloid volume, and the highest rocuronium and vasopressor exposure. This cluster also exhibited enrichment for hepatic disease, transplantation, and end-stage renal disease. These findings demonstrate that pre-operative PPG waveform phenotyping can reveal latent physiologic risk signatures with meaningful intraoperative implications. Future work will evaluate prospective perioperative risk assessment and decision-support workflows.

Speaker(s):
Aditya Nagori, PhD
Duke

Author(s):
Ayush Gautam, Bachelors - Duke University; Junseob Kim, MS - Duke; Ayman Ali, MD - Duke School of Medicine; Rishikesan Kamaleswaran, PhD - Duke School of Medicine;

---

Evaluating Representation Embeddings from LLMs and Time-Series Foundation Models for Wearable Accelerometer-Based Health Prediction

Presentation Type: Paper - Student
Presentation Time: 02:12 PM - 02:24 PM

Primary Track: Data Science/Artificial Intelligence


Wearable accelerometer data capture rich behavioral signals relevant for personalized health, yet the comparative evidence on modern representation-learning approaches remains limited. Using NHANES accelerometer data, we evaluated three representation families for predicting multiple clinical outcomes: simple entropy-based features, pretrained large-language-model (LLM) embeddings, and time-series foundation model embeddings. Outcomes included overweight status, lipid biomarkers, glucose, arthritis, and cancers. Across all endpoints, entropy-based features consistently matched or outperformed embedding approaches. LLM-derived embeddings offered only marginal improvements (ΔAUC≈0.01-0.05), and time-series foundation model embeddings provided minimal value regardless of sequence resolution. Prompt-based LLM reasoning performed worst (AUC≈0.56-0.65), demonstrating limited ability to infer quantitative physiological states from structured text. These results highlight the strength of simple variability features and underscore the need for domain-aligned pretraining in future time-series foundation models for health sensing.

Speaker(s):
Soomin You, MS
Columbia University

Author(s):
Soomin You, MS - Columbia University; Tian Gu, PhD - Columbia University;

---

Long COVID Phenotype Risk Score Reveals Decreased Steps and Elevated Resting Heart Rate Before and After Infection

Presentation Type: Podium Abstract
Presentation Time: 02:24 PM - 02:36 PM

Primary Track: Clinical Research Informatics


Long COVID remains difficult to identify in EHR data. We constructed a phenotype risk score (PheRS) using incident PheWAS in All of Us, compared it with the RECOVER computable phenotype and LATCH, and analyzed wearable data in PheRS-classified cases. The PheRS balanced sensitivity and specificity better than existing approaches. Those with long COVID showed lower step counts and higher resting heart rate pre-infection, but only post-infection step decline persisted relative to individual baselines.

Speaker(s):
Bennett Waxse, MD, PhD
NIAID/CNH

Author(s):
Bennett Waxse, MD, PhD - NIAID/CNH; Evelynne Fulda, MSc - NIH/University of Oxford; Joshua Denny, MD, MS - National Institutes of Health;

---

Temporal Recurrent Neural Networks for Predicting Acute Kidney Injury Recovery by Time of Discharge

Presentation Type: Paper - Regular
Presentation Time: 02:36 PM - 02:48 PM

Primary Track: Data Science/Artificial Intelligence


Acute Kidney Injury (AKI) is a common complication in hospitalized patients and is associated with increased in-hospital mortality, readmission, and chronic kidney disease. Early identification of patients at risk of AKI non-recovery can improve discharge planning and follow-up. Using a retrospective cohort of 7,667 patient encounters diagnosed with AKI from the University of California San Diego Health, we compared traditional machine learning (ML) and temporal deep learning (DL) models to predict three AKI recovery outcomes: Recovery, Partial Recovery, and Non-Recovery. The ML models evaluated were Logistic Regression, Random Forest, and XGBoost; while, the DL models were Gated Recurrent Unit (GRU) and Long-Short Term Memory. In the test set, DL models consistently outperformed traditional ML approaches. The GRU model achieved the highest Macro-Area Under the Curve (AUC) (0.822) with strong discrimination for the Non-Recovery class (AUC 0.932). This work demonstrates that temporal modeling of clinical trajectories can enhance AKI recovery prediction.

Speaker(s):
Nathan Tran, BS
UC San Diego

Author(s):
Nathan Tran, BS - UC San Diego; Zaid YOUSIF, PharmD, MAS - UCSD; Ambarish Athavale, MD - UCSD; Shamim Nemati, PhD - UCSD; Etienne Macedo, MD, PhD, FASN - UCSD;

---

Digital Health Technology Burden and Frustration Among Patients with Multimorbidity

Presentation Type: Paper - Regular
Presentation Time: 02:48 PM - 03:00 PM

Primary Track: Clinical Research Informatics


This study aimed to characterize patterns of digital health technology (DHT) use and examine the relationships between multimorbidity, DHT adoption, and user-reported frustration, with a focus on identifying socioeconomic and phenotypic determinants of digital health burden. A cross-sectional analysis was conducted using nationally representative public data from HINTS 7 (2024). Adults with at least one chronic condition were included (N=3,753). The outcomes were the number of DHTs used and frustration with digital tools. Multivariable logistic regression models were employed, adjusting for sociodemographic and clinical variables. Phenotypic subgroup analyses were conducted based on multimorbidity, DHT use, and frustration patterns.Among 3,753 participants, 46.9% had multimorbidity. Participants with multimorbidity reported using a greater number of DHTs on average (mean = 3.9 vs. 3.6, p < 0.001) compared to those with a single condition, yet exhibited significantly higher rates of frustration with digital tasks (61.2% vs. 54.1%, p < 0.001). Both higher income and educational groups were associated with lower odds of frustration. Greater DHT use was also independently associated with reduced frustration (OR = 0.78, 95% CI: 0.71-0.87, p < 0.01). Phenotypic subgroup analysis further identified individuals with multimorbidity and low DHT use as the most vulnerable profile, characterized by older age, lower socioeconomic status, and the highest frustration prevalence (60.3%).While individuals with multimorbidity use more DHTs, they experience greater frustration, particularly those with lower socioeconomic status. However, higher engagement with DHTs is associated with lower frustration, suggesting that technology proficiency may mitigate burden. Targeted interventions addressing digital literacy and user-centered.

Speaker(s):
Jiancheng Ye, PhD
Weill Cornell Medicine

Author(s):
Haoxin Chen, MS - The University of Hong Kong; Jiancheng Ye, PhD - Weill Cornell Medicine;

---

Characterizing Emergency Department Care Trajectories and Post-Acute Utilization in Hypertensive Patients

Presentation Type: Podium Abstract
Presentation Time: 03:00 PM - 03:12 PM

Primary Track: Clinical Research Informatics


Hypertension affects nearly half of the adult population in the United States and is a primary or contributing cause of over 690,000 deaths annually. With U.S. Emergency Departments (EDs) managing approximately 155 million visits per year, hypertension is frequently encountered not as an isolated clinical emergency but as a prevalent comorbidity that fundamentally complicates patient flow and resource utilization. Current estimates indicate that over 25% of ED patients have underlying hypertension, yet fewer than 30% receive successful referrals for outpatient follow-up, revealing a critical disconnect in care continuity. While existing literature has primarily focused on acute hypertensive crisis management, the broader impact of hypertension on macroscopic care pathways—from initial ED disposition to post-acute care transitions—remains inadequately characterized. To address this knowledge gap, this study leverages the MIMIC-IV database to characterize distinct care trajectories and downstream healthcare utilization patterns among hypertensive versus non-hypertensive patients, with the aim of supporting evidence-based workflow optimization and resource allocation.

Speaker(s):
Jiancheng Ye, PhD
Weill Cornell Medicine

Author(s):
Jiancheng Ye, PhD - Weill Cornell Medicine;

---