AI-Driven Healthcare: Bridging Systems Science and Clinical Practice

Permanent URI for this collectionhttps://hdl.handle.net/10125/112473

Browse

Now showing 1 - 12 of 12

Predicting Pediatric Surgical Case Duration Using Machine Learning: Leveraging Team Dynamics and Operational Features
(2026-01-06) Adams, Katherine; Yang, Yang; Garcia, Gian-Gabriel; Roy, Arka; Leong, Xin-Li; Chung, Dai; Ahuja, Vishal
Accurate surgical case duration prediction is critical for optimizing pediatric healthcare operations. We developed machine learning models (ML) to predict pediatric surgical case duration using novel non-clinical features and compared their performance to existing scheduling estimates. Using 202,149 surgical records from Children's Health Dallas, we incorporated features related to team familiarity, surgeon experience, and operational context alongside clinical variables. Among the ML models considered, LightGBM performed best, reducing median prediction error from −12.00 to −0.69 minutes. Five of the ten most important features were non-clinical, highlighting operational factors' significance. Performance gains were greatest in Otolaryngology (19.9%), Gastroenterology (19.8%), and Orthopedics (30.4%), and for patients aged 2–9 years (20.8%). These findings demonstrate that incorporating team dynamics and operational factors into ML models may significantly improve surgical duration predictions, supporting more accurate pediatric scheduling and potentially saving $407–$701 per case in operating room costs.
A Healthcare-Specific Classification Schema for Conversational AI Agents
(2026-01-06) Sriram, Karthikeyan
Recent advances in natural language processing (NLP) have accelerated the development of conversational agents (chatbots) in healthcare. Despite widespread adoption, a domain-specific classification schema tailored to the healthcare context is lacking. This paper proposes a novel classification framework for healthcare chatbots based on three dimensions: (1) the patient's journey within the healthcare system, (2) the chatbot’s functional purpose, and (3) the end-user role. Using morphological analysis and literature-based review, we categorize existing healthcare chatbot implementations within this schema. The framework reveals adoption trends, design implications, and gaps for future research, offering a structured basis for more effective and targeted chatbot development in healthcare.
PatientLens: AI-Enabled Interactive Avatars for Patient Report Summarization and Visualization
(2026-01-06) Singh, Utkarsh; Danzy Iii, Bretho; Umar, Afzal; Naqvi, Syed; Abdul-Muhsin, Haider; Cobran, Ewan; Riaz, Irbaz; Bryan, Chris
In clinical environments, doctors often must review large amounts of patient reports prior to consults and check-ups --- a task that is time-consuming, cognitively taxing, and prone to errors. We investigate how to improve this workflow via the use of AI-driven virtual avatars that enable clinicians to query, and summarize information from text-based patient reports. While the use of AI (specifically LLMs) brings significant potential benefits for clinical settings, it also presents critical challenges such as hallucination. Based on robust discussions and iterative prototyping with clinicians, we develop a human-in-the-loop approach that supports interactively creating and refining virtual avatars that visually present patient information while efficiently supporting LLM oversight and transparency. Evaluations help validate that the developed tool, nicknamed PatientLens, supports clinical review and summarization workflows. We also discuss how lessons learned during this project can enhance healthcare communication, optimize clinical workflows, and support improved health equity and outcomes.
Interpretable/Explainable Predictive Modeling with Perioperative Dataset
(2026-01-06) Dutta, Anandi; Li, Chih-Ying; Mruthyunjaya, Shivani
This study focuses on analyzing a perioperative dataset to extract knowledge and gain valuable insights for developing predictive models that enable real-time risk assessment and early intervention strategies for patients. We performed our analysis on several fronts, including K-means Clustering, Association Rule Mining, and SHAP analysis. Moreover, we developed two deep Learning models based on MambaNet and TabNet, achieving high AUROC scores of 0.946 and 0.923, respectively. The model’s objective is to predict 30-day mortality after surgery.
Bridging Science and Medicine with the ChRIS Research Integration System
(2026-01-06) Zhang, Jennings; Pienaar, Rudolph
ML, AI, and cloud computing revolutionize medical imaging research, yet the lag on the clinical side in technological advancement limits the direct impact of research innovation can have on patient care. We present ChRIS, a software platform and science gateway for deployment of computational research applications across various environments, including clinical settings. It provides an end-to-end solution for research and clinical workflows, starting with DICOM data retrieval from a hospital PACS to running analysis pipelines on Kubernetes or HPC. Using ChRIS, our research center rapidly develops and deploys advanced medical software following iterative development timelines.
Realistic Counterfactual Explanations for Clinical AI-decision Aid on Computed Tomography for Adaptive Radiotherapy
(2026-01-06) Heising, Luca; Ou, Carol; Verhaegen, Frank; Wolfs, Cecile; Cinquini, Martina; Spinnato, Francesco; Jacobs, Maria; Guidotti, Riccardo
Adaptive radiotherapy (ART) aims to detect changing patient anatomy and adapt the radiotherapy (RT) treatment plan accordingly, improving accuracy and outcomes. Artificial intelligence (AI) is introduced into this workflow to facilitate adaptation and manage the increased workload. However, the complexity and opacity of AI models used in ART pose challenges for clinical adoption. Outputs of traditional explainable AI methods, such as heatmaps, often lack consistency and can induce cognitive bias in clinicians' decision-making. This study introduces a novel approach using counterfactual explanations and domain knowledge to enhance transparency and trust in ART systems. By leveraging the physical properties of computed tomography images the proposed method generates realistic and semantically meaningful counterfactuals. These explanations help clinicians understand the clinical relevance of AI-detected deviations in the treatment, supporting safer and more effective decision-making for RT.
Deep Learning vs. Novice User for Needle Tip Tracking in Ultrasound-Guided Intravenous Access
(2026-01-06) Walsh, Carrie; Chu , David; Goldsmith, Andrew; Fischetti, Chanel; Driver, Lachlan; Taub, Olivia Ann; Blaivas, Laura; Duggan, Nicole M.; Blaivas, Michael; Shokoohi, Hamid
Background:. Ultrasound guided peripheral intravenous (US-pIV) access can reduce patient morbidity and mortality and improve the quality of care. Currently, there are no out-of-plane artificial intelligence tools designed to help learners reach competency with this procedure. Methods: We developed a deep learning (DL) model to identify needle-tip tracking during simulated US-pIV placement with out-of-plane technique. The sensitivity and specificity of this DL model was compared to a novice participant group. Results: Our DL model outperformed our novice group with an increased accuracy score (0.89 vs. 0.82 (95% CI 0.79–0.85), and higher sensitivity (0.91 vs. 0.82 (95% CI 0.79–0.85) and specificity scores (0.85 vs. 0.81 (95% CI 0.74–0.86), although these differences were not statistically significant. Conclusion: DL has the potential to enhance the learning curve associated with performing US-pIV access to ensure safer and more efficient outcomes for patients.
AI-Driven Prediction of ROSC Using EtCO2 Trends and Response Dynamics Following Vasopressor Administration
(2026-01-06) Raya Krishnamoorthy, Banu Priya; Nassal, Michelle; Wang, Henry; Elola, Andoni; Aramendi, Elisabete; Jaureguibeitia, Xabier; Idris, Ahamed; Panchal, Ashish; Ulintz, Alexander; Sugavanam, Nithin; Ertin, Emre
Capnography is widely used during resuscitation to monitor ventilation and perfusion. Prior studies have shown that out-of-hospital cardiac arrest (OHCA) outcomes are associated with end-tidal capnography (EtCO2) trends. However, the use of artificial intelligence (AI) to predict return of spontaneous circulation (ROSC) based on EtCO2 trends remains unexplored. We aimed to develop an AI model to predict ROSC using EtCO2 trends and identify influential features. We conducted a secondary analysis of the Pragmatic Airway Resuscitation Trial (PART), including previously calculated EtCO2 slope and post-vasopressor response features, along with patient demographics. An AI-based Random Forest classifier was trained to predict ROSC, and feature importance scores were extracted. The model achieved 83% accuracy, indicating strong performance in identifying both ROSC and non-ROSC cases. Of the top five features ranked by the model, four were related to EtCO2 trends and response to vasopressor administration. AI-driven EtCO2 algorithms may strongly influence resuscitation decisions.
AI Enabled Content Analysis of Peer vs. Professional Telephonic Coaching to Support Online Mindfulness Based Cognitive Therapy for Prevention of Perinatal Depression Recurrence
(2026-01-06) Beck, Arne
Online Mindfulness-Based Cognitive Therapy (MBCT) is effective in preventing perinatal depression recurrence. However, telephonic coaching support using Motivational Interviewing (MI) is important to foster program engagement. Utilizing peers with lived experience of perinatal depression as coaches may expand the mental health workforce and offer unique benefits over mental health professionals. This study used an AI program (Lyssn) to compare coaching calls from trained peers vs. mental health professionals. Transcribed calls with 209 participants randomized to receive coaching from peers and 214 to receive coaching from mental health professionals were coded for MI proficiency and percent of MI-type statements used. Among 682 calls, Professional coaches had significantly higher MI proficiency ratings than peer coaches whereas peer coaches had a higher percent of speech coded as giving information that included self-disclosure and psycho-ed style interventions. However, program engagement and clinical outcomes were equivalent, suggesting that different coaching styles may be equally effective.
FedSDMU: A new Paradigm for Healthcare Data Privacy Compliance using Federated Synthesis, Differential Privacy, and Machine Unlearning
(2026-01-06) Yadav, Parul; D'Souza, Ivan; Mishra, Sushma
This study introduces FedSDMU, an innovative framework integrating Federated Learning (FL), Synthetic Data Generation (SDG), Differential Privacy (DP), and Machine Unlearning (MU) to enhance privacy-preserving healthcare analytics. FedSDMU enables collaborative model training across decentralized datasets without sharing raw data, safeguarding patient privacy. It strengthens security by injecting noise during training, preventing individual data inference. A key innovation is its federated synthesis, leveraging DP for secure generative model training and MU for data deletion, allowing patients to revoke consent while ensuring regulatory compliance. This framework outperforms existing models, strictly aligning with HIPAA and GDPR, while mitigating risks, enhancing prediction quality, addressing bias, and refining models by securely deleting data points. By tackling data scarcity, imbalance, and heterogeneity, FedSDMU establishes a robust and reliable analytical system. Its contributions pave the way for future studies, extending the use of privacy-preserving AI applications in healthcare.
Calibrating Trust in AI for Clinical Dentistry: A Qualitative Study of Human- AI Collaboration and Adoption Dynamics
(2026-01-06) Nguyen, Thomas; Pinsonneault, Alain
Artificial intelligence (AI) is increasingly integrated into clinical dentistry, offering enhancements in diagnostic accuracy, workflow efficiency, and decision-making support. However, its adoption remains uneven, largely due to challenges in trust calibration. This involves finding the right balance between overreliance and underuse. This qualitative study investigates how dental professionals perceive, develop, and recalibrate trust in AI tools, drawing on seven in-depth interviews with clinicians across specialties and geographies. Findings reveal three distinct user clusters: Cautious Verifiers, Balanced Validators, and Overreliant New Graduates, each with unique trust behaviors and verification strategies. Key factors influencing trust include interpretability, reliability, workflow integration, and clinician education. The study introduces an empirically grounded adaptation of the Inverted U-Model of AI Utilization, which links optimal performance to moderate verification (20–50% of AI outputs). These insights inform a trust-centered framework for safe and effective AI adoption, with implications for system design, training, and regulatory policy in dentistry and beyond.
Introduction to the Minitrack on AI-Driven Healthcare: Bridging Systems Science and Clinical Practice
(2026-01-06) Ahmadi, Farzin; Jambaulikar, Guruprasad; Vaz, Clint

Browse

Recent Submissions