Pavitra Krishnaswamy
Research Subjects
Pavitra Krishnaswamy has a total of 113 co-authors. Here are the top co-authors by number of shared publications:
- Nancy F. Chen
Agency for Science, Technology and Research (A*STAR), Singapore, Singapore
6 shared publications - Mojisola Erdt
Wee Kim Wee School of Communication and Information, Nanyang Technological University, Singapore, Singapore
4 shared publications - Yu En Chan
Institute for Infocomm Research (I2R), A*STAR, Singapore, Singapore
4 shared publications - Angela Ng
Changi General Hospital, Singapore, Singapore
3 shared publications - Bryan Peide Choo
Changi General Hospital, Singapore, Singapore
3 shared publications - Geok Choo Lim
Changi General Hospital, Singapore, Singapore
3 shared publications
Latest Publications
- Open Access
- 23 October 2026
- Published by JMIR Publications Inc. in Journal of Medical Internet Research
- Vol. 28, e91058
Evaluation of a Clinical Protocol–Guided Chatbot to Support Telecarers in Diabetes Management: Embedded Mixed Methods Study
Background: Type 2 diabetes mellitus (T2DM) is a chronic condition that requires ongoing self-management. While chatbots offer scalable monitoring and management, evidence on their integration into structured hospital tele-support workflows, where patient engagement and digital readiness may differ from those in standalone digital interventions, remains limited. Objective: This study aimed to describe the design of a clinical protocol–guided chatbot, Sharing Information with Users on Diabetes Actions and Interventions ( Siu Dai ), and evaluate its feasibility, usability, and acceptability among patients and telecarers within a hospital tele-support program. Methods: We used an embedded mixed methods design (QUAL [quan]) guided by the World Health Organization (WHO) monitoring and evaluating digital health interventions and the theoretical framework of acceptability. Through a co-design process involving clinicians, researchers, telecarers, and developers, the chatbot was aligned with established clinical protocols. It featured 7 precall sessions designed to collect patient data, generate summary reports for staff, and deliver targeted educational content. Patients (≥21-year-old individuals with T2DM) and telecarer nurses (with experience ≥6 mo) were recruited from an existing tele-support program. Data sources included interviews, online surveys (5-point Likert scale), and chatbot usage logs. Qualitative data were analyzed thematically; quantitative data were summarized descriptively to supplement and contextualize the qualitative findings. Results: Fourteen participants (7 patients and 7 telecarers) completed both interviews and surveys and were included in the analysis. Patients (mean age 45.3, SD 15.4 y) had a median diabetes duration of 10 (IQR 3-25; range 1‐30) years, and all were digitally literate. Telecarers had a median of 5 (IQR 4-9) years of telehealth experience. Quantitatively, patients showed high engagement, completing 79% of started sessions with a median completion time of 2 minutes and 58 seconds. Patient satisfaction was high, with 86% (6/7) willing to recommend the tool; however, telecarer endorsement for universal rollout to patients in other clinical programs or outpatient clinics was more cautious (3/7, 43%). Qualitative findings revealed 5 themes: (1) perceptions of chatbot usability, (2) alignment and integration into telecarers’ processes, (3) factors influencing acceptability of the chatbot, (4) perceived efficacy of the chatbot for T2DM self-management and care delivery, and (5) suggested improvements and enhancements. Integrated findings revealed that the chatbot was usable and functioned as a precall support tool that optimized the personalization of tele-support sessions. However, its perceived efficacy in self-management behavior was limited by this cohort’s extensive experience with T2DM management. Key barriers to adoption included administrative burden for telecarers due to a lack of electronic medical record integration and content redundancy for patients. Suggested enhancements included multilingual support and adaptive content catering to different levels of disease-specific literacy. Conclusions: A clinical protocol–guided chatbot is feasible for hospital tele-support, complementing rather than replacing human care. It automates routine assessments and reinforces education, but future iterations must prioritize workflow integration and adaptive content to sustain engagement.
- 23 October 2026
- Published by JMIR Publications Inc. in Journal of Medical Internet Research
- Vol. 28, e91058
Open AccessEvaluation of a Clinical Protocol–Guided Chatbot to Support Telecarers in Diabetes Management: Embedded Mixed Methods Study
Background: Type 2 diabetes mellitus (T2DM) is a chronic condition that requires ongoing self-management. While chatbots offer scalable monitoring and management, evidence on their integration into structured hospital tele-support workflows, where patient engagement and digital readiness may differ from those in standalone digital interventions, remains limited. Objective: This study aimed to describe the design of a clinical protocol–guided chatbot, Sharing Information with Users on Diabetes Actions and Interventions ( Siu Dai ), and evaluate its feasibility, usability, and acceptability among patients and telecarers within a hospital tele-support program. Methods: We used an embedded mixed methods design (QUAL [quan]) guided by the World Health Organization (WHO) monitoring and evaluating digital health interventions and the theoretical framework of acceptability. Through a co-design process involving clinicians, researchers, telecarers, and developers, the chatbot was aligned with established clinical protocols. It featured 7 precall sessions designed to collect patient data, generate summary reports for staff, and deliver targeted educational content. Patients (≥21-year-old individuals with T2DM) and telecarer nurses (with experience ≥6 mo) were recruited from an existing tele-support program. Data sources included interviews, online surveys (5-point Likert scale), and chatbot usage logs. Qualitative data were analyzed thematically; quantitative data were summarized descriptively to supplement and contextualize the qualitative findings. Results: Fourteen participants (7 patients and 7 telecarers) completed both interviews and surveys and were included in the analysis. Patients (mean age 45.3, SD 15.4 y) had a median diabetes duration of 10 (IQR 3-25; range 1‐30) years, and all were digitally literate. Telecarers had a median of 5 (IQR 4-9) years of telehealth experience. Quantitatively, patients showed high engagement, completing 79% of started sessions with a median completion time of 2 minutes and 58 seconds. Patient satisfaction was high, with 86% (6/7) willing to recommend the tool; however, telecarer endorsement for universal rollout to patients in other clinical programs or outpatient clinics was more cautious (3/7, 43%). Qualitative findings revealed 5 themes: (1) perceptions of chatbot usability, (2) alignment and integration into telecarers’ processes, (3) factors influencing acceptability of the chatbot, (4) perceived efficacy of the chatbot for T2DM self-management and care delivery, and (5) suggested improvements and enhancements. Integrated findings revealed that the chatbot was usable and functioned as a precall support tool that optimized the personalization of tele-support sessions. However, its perceived efficacy in self-management behavior was limited by this cohort’s extensive experience with T2DM management. Key barriers to adoption included administrative burden for telecarers due to a lack of electronic medical record integration and content redundancy for patients. Suggested enhancements included multilingual support and adaptive content catering to different levels of disease-specific literacy. Conclusions: A clinical protocol–guided chatbot is feasible for hospital tele-support, complementing rather than replacing human care. It automates routine assessments and reinforces education, but future iterations must prioritize workflow integration and adaptive content to sustain engagement.
- Open Access
- 4 September 2026
- Published by JMIR Publications Inc. in Journal of Medical Internet Research
- Vol. 28, e86629
Graph-Based Modeling of Behavioral Transitions for Digital Health Applications: Trajectory Analysis Framework
Background: Digital health applications generate rich behavioral data; yet, how users transition between behavioral states remains poorly understood. Existing approaches show limited capacity to capture and represent the evolving and nuanced nature of real-world behavioral dynamics, which are essential for informing personalized behavior change interventions. Objective: This study aimed to understand how users’ behaviors evolve within digital health applications by developing a data-driven approach that captures transition dynamics across behavioral features. We aimed to validate modeled transitions against observed user data, analyze transition pathways, and derive interpretable insights. Methods: We analyzed 32 weeks of data from 36,574 users from a population health program run by the Health Promotion Board (HPB) in Singapore. We developed a graph-based behavioral trajectory model (GraphBeTraM ) that models behavioral transitions as shortest paths through user similarity graphs. By aggregating paths over time, the model captures users’ progression toward target behavioral states and quantifies the direction, magnitude, and timing of change. We validated modeled transitions against real-world data by comparing (1) transition matrices and (2) feature-level pathways, using Spearman correlation, Fisher z-transformed means, and cosine distance. We focused on high-variance features and used heatmaps to visualize patterns of change. Finally, characterized transition pathways between behavioral states and distinguished generalizable and context-specific features. Results: We observed strong alignment between observed and modeled transition matrices with strong correlation (Spearman ρ=0.82; P <.001) and low cosine distance (0.05), suggesting that real-world behavioral transitions can be effectively represented through shortest paths with GraphBeTraM. Feature-level pathway validation showed consistent patterns of change across both personalized and real-world pathways. Physical activity features like weekly moderate to vigorous physical activity (MVPA) emerged as stable, generalizable signals of change at the population level, with strong mean correlation (ρ_ personalized =0.97; ρ_ real-world =0.79) and low mean cosine distance (cosine_ personalized =0.24; cosine_ real-world =0.41) across both personalized and real-world transitions. In contrast, features related to personal preferences, including time-related activity patterns (eg, proportion of weekday to weekend MVPA), purchase preferences for healthy foods and drinks, and engagement indicators (eg, last contact with the program or app), exhibited context-specific relevance and reflected change at an individual level. An in-depth transition analysis from an active state with prolonged sedentary periods to an active state with healthy eating habits revealed interpretable patterns in onset, magnitude, and rate of change. Specifically, features such as preferences for healthy food and drink purchases emerged later in the trajectory but then showed sharper and faster transitions once these behaviors began to shift. These insights highlight how GraphBeTraM can guide nudging strategies in alignment with natural behavior dynamics, including both stable and later-onset patterns. Conclusions: GraphBeTraM provides an interpretable framework for modeling behavioral transitions in digital health applications and captures key dynamics that support personalized intervention design.
- 4 September 2026
- Published by JMIR Publications Inc. in Journal of Medical Internet Research
- Vol. 28, e86629
Open AccessGraph-Based Modeling of Behavioral Transitions for Digital Health Applications: Trajectory Analysis Framework
Background: Digital health applications generate rich behavioral data; yet, how users transition between behavioral states remains poorly understood. Existing approaches show limited capacity to capture and represent the evolving and nuanced nature of real-world behavioral dynamics, which are essential for informing personalized behavior change interventions. Objective: This study aimed to understand how users’ behaviors evolve within digital health applications by developing a data-driven approach that captures transition dynamics across behavioral features. We aimed to validate modeled transitions against observed user data, analyze transition pathways, and derive interpretable insights. Methods: We analyzed 32 weeks of data from 36,574 users from a population health program run by the Health Promotion Board (HPB) in Singapore. We developed a graph-based behavioral trajectory model (GraphBeTraM ) that models behavioral transitions as shortest paths through user similarity graphs. By aggregating paths over time, the model captures users’ progression toward target behavioral states and quantifies the direction, magnitude, and timing of change. We validated modeled transitions against real-world data by comparing (1) transition matrices and (2) feature-level pathways, using Spearman correlation, Fisher z-transformed means, and cosine distance. We focused on high-variance features and used heatmaps to visualize patterns of change. Finally, characterized transition pathways between behavioral states and distinguished generalizable and context-specific features. Results: We observed strong alignment between observed and modeled transition matrices with strong correlation (Spearman ρ=0.82; P <.001) and low cosine distance (0.05), suggesting that real-world behavioral transitions can be effectively represented through shortest paths with GraphBeTraM. Feature-level pathway validation showed consistent patterns of change across both personalized and real-world pathways. Physical activity features like weekly moderate to vigorous physical activity (MVPA) emerged as stable, generalizable signals of change at the population level, with strong mean correlation (ρ_ personalized =0.97; ρ_ real-world =0.79) and low mean cosine distance (cosine_ personalized =0.24; cosine_ real-world =0.41) across both personalized and real-world transitions. In contrast, features related to personal preferences, including time-related activity patterns (eg, proportion of weekday to weekend MVPA), purchase preferences for healthy foods and drinks, and engagement indicators (eg, last contact with the program or app), exhibited context-specific relevance and reflected change at an individual level. An in-depth transition analysis from an active state with prolonged sedentary periods to an active state with healthy eating habits revealed interpretable patterns in onset, magnitude, and rate of change. Specifically, features such as preferences for healthy food and drink purchases emerged later in the trajectory but then showed sharper and faster transitions once these behaviors began to shift. These insights highlight how GraphBeTraM can guide nudging strategies in alignment with natural behavior dynamics, including both stable and later-onset patterns. Conclusions: GraphBeTraM provides an interpretable framework for modeling behavioral transitions in digital health applications and captures key dynamics that support personalized intervention design.
- 1 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Biomedical and Health Informatics
- Vol. PP (21682194) , 1-14
Towards Reliable Prediction: A Bayesian Continual Learning Approach for Clinical Time-series Data
Deep learning models are increasingly used for making predictions based on clinical time series data, but model generalization remains a challenge. Continual learning approaches, which preserve representations while learning new distributions, are suitable for addressing this challenge. We propose Continual Bayesian Long Short Term Memory (C-BLSTM), a continual learning algorithm based on the Bayesian LSTM model for domain incremental learning. C-BLSTM continually learns a sequence of tasks by combining architectural pruning, variational inference-based regularization, and coreset replay strategies. In extensive experiments on two public electronic medical record datasets for mortality prediction, we show that C-BLSTM outperforms many state-of-the-art continual learning approaches. Further, we apply the C-BLSTM to two realworld clinical time series datasets for prediction of readmission risk in patients with heart failure and glycated haemoglobin outcomes in patients with type 2 diabetes.First, we show that these datasets exhibit domain incremental characteristics with significant drifts in their marginal distributions and moderate drifts in their conditional distributions. Then, we demonstrate that the C-BLSTM improves generalization in five diverse realworld scenarios spanning temporal, site, device, case mix, and ethnicity shifts, both in terms of performance and reliability of predictions.
- 1 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Biomedical and Health Informatics
- Vol. PP (21682194) , 1-14
Towards Reliable Prediction: A Bayesian Continual Learning Approach for Clinical Time-series Data
Deep learning models are increasingly used for making predictions based on clinical time series data, but model generalization remains a challenge. Continual learning approaches, which preserve representations while learning new distributions, are suitable for addressing this challenge. We propose Continual Bayesian Long Short Term Memory (C-BLSTM), a continual learning algorithm based on the Bayesian LSTM model for domain incremental learning. C-BLSTM continually learns a sequence of tasks by combining architectural pruning, variational inference-based regularization, and coreset replay strategies. In extensive experiments on two public electronic medical record datasets for mortality prediction, we show that C-BLSTM outperforms many state-of-the-art continual learning approaches. Further, we apply the C-BLSTM to two realworld clinical time series datasets for prediction of readmission risk in patients with heart failure and glycated haemoglobin outcomes in patients with type 2 diabetes.First, we show that these datasets exhibit domain incremental characteristics with significant drifts in their marginal distributions and moderate drifts in their conditional distributions. Then, we demonstrate that the C-BLSTM improves generalization in five diverse realworld scenarios spanning temporal, site, device, case mix, and ethnicity shifts, both in terms of performance and reliability of predictions.
- Open Access
- 30 October 2024
- Published by JMIR Publications Inc. in Journal of Medical Internet Research
- Vol. 26 (1) , e46983
Characterization of Telecare Conversations on Lifestyle Management and Their Relation to Health Care Utilization for Patients with Heart Failure: Mixed Methods Study
Background: Telehealth interventions where providers offer support and coaching to patients with chronic conditions such as heart failure (HF) and type 2 diabetes mellitus (T2DM) are effective in improving health outcomes. However, the understanding of the content and structure of these interactions and how they relate to health care utilization remains incomplete. Objective: This study aimed to characterize the content and structure of telecare conversations on lifestyle management for patients with HF and investigate how these conversations relate to health care utilization. Methods: We leveraged real-world data from 50 patients with HF enrolled in a postdischarge telehealth program, with the primary intervention comprising a series of telephone calls from nurse telecarers over a 12-month period. For the full cohort, we transcribed 729 English-language calls and annotated conversation topics. For a subcohort (25 patients with both HF and T2DM), we annotated lifestyle management content with fine-grained dialogue acts describing typical conversational structures. For each patient, we identified calls with unusually high ratios of utterances on lifestyle management as lifestyle-focused calls. We further extracted structured data for inpatient admissions from 6 months before to 6 months after the intervention period. First, to understand conversational structures and content of lifestyle-focused calls, we compared the number of utterances, dialogue acts, and symptom attributes in lifestyle-focused calls to those in calls containing but not focused on lifestyle management. Second, to understand the perspectives of nurse telecarers on these calls, we conducted an expert evaluation where 2 nurse telecarers judged levels of concern and follow-up actions for lifestyle-focused and other calls (not focused on lifestyle management content). Finally, we assessed how the number of lifestyle-focused calls relates to the number of admissions, and to the average length of stay per admission. Results: In comparative analyses, lifestyle-focused calls had significantly fewer utterances (P=.01) and more dialogue acts (Padj=.005) than calls containing but not focused on lifestyle management. Lifestyle-focused calls did not contain deeper discussions on clinical symptoms. These findings indicate that lifestyle-focused calls entail short, intense discussions with greater emphasis on understanding patient experience and coaching than on clinical content. In the expert evaluation, nurse telecarers identified 24.2% (29/120) of calls assessed as concerning enough for follow-up. For these 29 calls, nurse telecarers were more attuned to concerns about symptoms and vitals (19/29, 65.5%) than lifestyle management concerns (4/29, 13.8%). The number of lifestyle-focused calls a patient had was modestly (but not significantly) associated with a lower average length of stay for inpatient admissions (Spearman ρ=-0.30; Padj=.06), but not with the number of admissions (Spearman ρ=-0.03; Padj=.84). Conclusions: Our approach and findings offer novel perspectives on the content, structure, and clinical associations of telehealth conversations on lifestyle management for patients with HF. Hence, our study could inform ways to enhance telehealth programs for self-care management in chronic conditions.
- 30 October 2024
- Published by JMIR Publications Inc. in Journal of Medical Internet Research
- Vol. 26 (1) , e46983
Open AccessCharacterization of Telecare Conversations on Lifestyle Management and Their Relation to Health Care Utilization for Patients with Heart Failure: Mixed Methods Study
Background: Telehealth interventions where providers offer support and coaching to patients with chronic conditions such as heart failure (HF) and type 2 diabetes mellitus (T2DM) are effective in improving health outcomes. However, the understanding of the content and structure of these interactions and how they relate to health care utilization remains incomplete. Objective: This study aimed to characterize the content and structure of telecare conversations on lifestyle management for patients with HF and investigate how these conversations relate to health care utilization. Methods: We leveraged real-world data from 50 patients with HF enrolled in a postdischarge telehealth program, with the primary intervention comprising a series of telephone calls from nurse telecarers over a 12-month period. For the full cohort, we transcribed 729 English-language calls and annotated conversation topics. For a subcohort (25 patients with both HF and T2DM), we annotated lifestyle management content with fine-grained dialogue acts describing typical conversational structures. For each patient, we identified calls with unusually high ratios of utterances on lifestyle management as lifestyle-focused calls. We further extracted structured data for inpatient admissions from 6 months before to 6 months after the intervention period. First, to understand conversational structures and content of lifestyle-focused calls, we compared the number of utterances, dialogue acts, and symptom attributes in lifestyle-focused calls to those in calls containing but not focused on lifestyle management. Second, to understand the perspectives of nurse telecarers on these calls, we conducted an expert evaluation where 2 nurse telecarers judged levels of concern and follow-up actions for lifestyle-focused and other calls (not focused on lifestyle management content). Finally, we assessed how the number of lifestyle-focused calls relates to the number of admissions, and to the average length of stay per admission. Results: In comparative analyses, lifestyle-focused calls had significantly fewer utterances (P=.01) and more dialogue acts (Padj=.005) than calls containing but not focused on lifestyle management. Lifestyle-focused calls did not contain deeper discussions on clinical symptoms. These findings indicate that lifestyle-focused calls entail short, intense discussions with greater emphasis on understanding patient experience and coaching than on clinical content. In the expert evaluation, nurse telecarers identified 24.2% (29/120) of calls assessed as concerning enough for follow-up. For these 29 calls, nurse telecarers were more attuned to concerns about symptoms and vitals (19/29, 65.5%) than lifestyle management concerns (4/29, 13.8%). The number of lifestyle-focused calls a patient had was modestly (but not significantly) associated with a lower average length of stay for inpatient admissions (Spearman ρ=-0.30; Padj=.06), but not with the number of admissions (Spearman ρ=-0.03; Padj=.84). Conclusions: Our approach and findings offer novel perspectives on the content, structure, and clinical associations of telehealth conversations on lifestyle management for patients with HF. Hence, our study could inform ways to enhance telehealth programs for self-care management in chronic conditions.
- Open Access
- 17 September 2024
- Published by Springer Nature in npj Digital Medicine
- Vol. 7 (1) , 1-12
A drug mix and dose decision algorithm for individualized type 2 diabetes management
Pharmacotherapy guidelines for type 2 diabetes (T2D) emphasize patient-centered care, but applying this approach effectively in outpatient practice remains challenging. Data-driven treatment optimization approaches could enhance individualized T2D management, but current approaches cannot account for drug-specific and dose-dependent variations in safety and efficacy. We developed and evaluated an AI Drug mix and dose Advisor (AIDA) for glycemic management, using electronic medical records from 107,854 T2D patients in the SingHealth Diabetes Registry. Given a patient’s medical profile, AIDA leverages a predict-then-optimize approach to identify the minimal drug mix and dose changes required to optimize glycemic control, subject to clinical knowledge-based guidelines. On unseen data from large internal, external, and temporal validation sets, AIDA recommendations were estimated to improve post-visit glycated hemoglobin (HbA1c) by an average of 0.40–0.68% over standard of care (P < 0.0001). In qualitative evaluations on 60 diverse cases by a panel of three endocrinologists, AIDA recommendations were mostly rated as reasonable and precise. Finally, AIDA’s ability to account for drug-dose specifics offered several advantages over competing methods, including greater consistency with practice preferences and clinical guidelines for practical but effective options, indication-based treatments, and renal dosing. As AIDA provides drug-dose recommendations to improve outcomes for individual T2D patients, it could be used for clinical decision support at point-of-care, especially in resource-limited settings.
- 17 September 2024
- Published by Springer Nature in npj Digital Medicine
- Vol. 7 (1) , 1-12
Open AccessA drug mix and dose decision algorithm for individualized type 2 diabetes management
Pharmacotherapy guidelines for type 2 diabetes (T2D) emphasize patient-centered care, but applying this approach effectively in outpatient practice remains challenging. Data-driven treatment optimization approaches could enhance individualized T2D management, but current approaches cannot account for drug-specific and dose-dependent variations in safety and efficacy. We developed and evaluated an AI Drug mix and dose Advisor (AIDA) for glycemic management, using electronic medical records from 107,854 T2D patients in the SingHealth Diabetes Registry. Given a patient’s medical profile, AIDA leverages a predict-then-optimize approach to identify the minimal drug mix and dose changes required to optimize glycemic control, subject to clinical knowledge-based guidelines. On unseen data from large internal, external, and temporal validation sets, AIDA recommendations were estimated to improve post-visit glycated hemoglobin (HbA1c) by an average of 0.40–0.68% over standard of care (P < 0.0001). In qualitative evaluations on 60 diverse cases by a panel of three endocrinologists, AIDA recommendations were mostly rated as reasonable and precise. Finally, AIDA’s ability to account for drug-dose specifics offered several advantages over competing methods, including greater consistency with practice preferences and clinical guidelines for practical but effective options, indication-based treatments, and renal dosing. As AIDA provides drug-dose recommendations to improve outcomes for individual T2D patients, it could be used for clinical decision support at point-of-care, especially in resource-limited settings.
- Open Access
- 12 January 2024
- Published by JMIR Publications Inc. in JMIR Preprints
Characterization of Telecare Conversations on Lifestyle Management and Their Relation to Health Care Utilization for Patients with Heart Failure: Mixed Methods Study (Preprint)
BACKGROUND Telehealth interventions where providers offer support and coaching to patients with chronic conditions such as heart failure (HF) and type 2 diabetes mellitus (T2DM) are effective in improving health outcomes. However, the understanding of the content and structure of these interactions and how they relate to health care utilization remains incomplete. OBJECTIVE This study aimed to characterize the content and structure of telecare conversations on lifestyle management for patients with HF and investigate how these conversations relate to health care utilization. METHODS We leveraged real-world data from 50 patients with HF enrolled in a postdischarge telehealth program, with the primary intervention comprising a series of telephone calls from nurse telecarers over a 12-month period. For the full cohort, we transcribed 729 English-language calls and annotated conversation topics. For a subcohort (25 patients with both HF and T2DM), we annotated lifestyle management content with fine-grained dialogue acts describing typical conversational structures. For each patient, we identified calls with unusually high ratios of utterances on lifestyle management as lifestyle-focused calls. We further extracted structured data for inpatient admissions from 6 months before to 6 months after the intervention period. First, to understand conversational structures and content of lifestyle-focused calls, we compared the number of utterances, dialogue acts, and symptom attributes in lifestyle-focused calls to those in calls containing but not focused on lifestyle management. Second, to understand the perspectives of nurse telecarers on these calls, we conducted an expert evaluation where 2 nurse telecarers judged levels of concern and follow-up actions for lifestyle-focused and other calls (not focused on lifestyle management content). Finally, we assessed how the number of lifestyle-focused calls relates to the number of admissions, and to the average length of stay per admission. RESULTS In comparative analyses, lifestyle-focused calls had significantly fewer utterances (P=.01) and more dialogue acts (Padj=.005) than calls containing but not focused on lifestyle management. Lifestyle-focused calls did not contain deeper discussions on clinical symptoms. These findings indicate that lifestyle-focused calls entail short, intense discussions with greater emphasis on understanding patient experience and coaching than on clinical content. In the expert evaluation, nurse telecarers identified 24.2% (29/120) of calls assessed as concerning enough for follow-up. For these 29 calls, nurse telecarers were more attuned to concerns about symptoms and vitals (19/29, 65.5%) than lifestyle management concerns (4/29, 13.8%). The number of lifestyle-focused calls a patient had was modestly (but not significantly) associated with a lower average length of stay for inpatient admissions (Spearman ρ=-0.30; Padj=.06), but not with the number of admissions (Spearman ρ=-0.03; Padj=.84). CONCLUSIONS Our approach and findings offer novel perspectives on the content, structure, and clinical associations of telehealth conversations on lifestyle management for patients with HF. Hence, our study could inform ways to enhance telehealth programs for self-care management in chronic conditions.
- 12 January 2024
- Published by JMIR Publications Inc. in JMIR Preprints
Open AccessCharacterization of Telecare Conversations on Lifestyle Management and Their Relation to Health Care Utilization for Patients with Heart Failure: Mixed Methods Study (Preprint)
BACKGROUND Telehealth interventions where providers offer support and coaching to patients with chronic conditions such as heart failure (HF) and type 2 diabetes mellitus (T2DM) are effective in improving health outcomes. However, the understanding of the content and structure of these interactions and how they relate to health care utilization remains incomplete. OBJECTIVE This study aimed to characterize the content and structure of telecare conversations on lifestyle management for patients with HF and investigate how these conversations relate to health care utilization. METHODS We leveraged real-world data from 50 patients with HF enrolled in a postdischarge telehealth program, with the primary intervention comprising a series of telephone calls from nurse telecarers over a 12-month period. For the full cohort, we transcribed 729 English-language calls and annotated conversation topics. For a subcohort (25 patients with both HF and T2DM), we annotated lifestyle management content with fine-grained dialogue acts describing typical conversational structures. For each patient, we identified calls with unusually high ratios of utterances on lifestyle management as lifestyle-focused calls. We further extracted structured data for inpatient admissions from 6 months before to 6 months after the intervention period. First, to understand conversational structures and content of lifestyle-focused calls, we compared the number of utterances, dialogue acts, and symptom attributes in lifestyle-focused calls to those in calls containing but not focused on lifestyle management. Second, to understand the perspectives of nurse telecarers on these calls, we conducted an expert evaluation where 2 nurse telecarers judged levels of concern and follow-up actions for lifestyle-focused and other calls (not focused on lifestyle management content). Finally, we assessed how the number of lifestyle-focused calls relates to the number of admissions, and to the average length of stay per admission. RESULTS In comparative analyses, lifestyle-focused calls had significantly fewer utterances (P=.01) and more dialogue acts (Padj=.005) than calls containing but not focused on lifestyle management. Lifestyle-focused calls did not contain deeper discussions on clinical symptoms. These findings indicate that lifestyle-focused calls entail short, intense discussions with greater emphasis on understanding patient experience and coaching than on clinical content. In the expert evaluation, nurse telecarers identified 24.2% (29/120) of calls assessed as concerning enough for follow-up. For these 29 calls, nurse telecarers were more attuned to concerns about symptoms and vitals (19/29, 65.5%) than lifestyle management concerns (4/29, 13.8%). The number of lifestyle-focused calls a patient had was modestly (but not significantly) associated with a lower average length of stay for inpatient admissions (Spearman ρ=-0.30; Padj=.06), but not with the number of admissions (Spearman ρ=-0.03; Padj=.84). CONCLUSIONS Our approach and findings offer novel perspectives on the content, structure, and clinical associations of telehealth conversations on lifestyle management for patients with HF. Hence, our study could inform ways to enhance telehealth programs for self-care management in chronic conditions.
- Open Access
- 4 August 2023
- Published by Association for Computing Machinery (ACM)
- p. 4673-4684
Deep Offline Reinforcement Learning for Real-world Treatment Optimization Applications
There is increasing interest in data-driven approaches for recommending optimal treatment strategies in many chronic disease management and critical care applications. Reinforcement learning methods are well-suited to this sequential decision-making problem, but must be trained and evaluated exclusively on retrospective medical record datasets as direct online exploration is unsafe and infeasible. Despite this requirement, the vast majority of treatment optimization studies use off-policy RL methods (e.g., Double Deep Q Networks (DDQN) or its variants) that are known to perform poorly in purely offline settings. Recent advances in offline RL, such as Conservative Q-Learning (CQL), offer a suitable alternative. But there remain challenges in adapting these approaches to real-world applications where suboptimal examples dominate the retrospective dataset and strict safety constraints need to be satisfied. In this work, we introduce a practical and theoretically grounded transition sampling approach to address action imbalance during offline RL training. We perform extensive experiments on two real-world tasks for diabetes and sepsis treatment optimization to compare performance of the proposed approach against prominent off-policy and offline RL baselines (DDQN and CQL). Across a range of principled and clinically relevant metrics, we show that our proposed approach enables substantial improvements in expected health outcomes and in consistency with relevant practice and safety guidelines.
- 4 August 2023
- Published by Association for Computing Machinery (ACM)
- p. 4673-4684
Open AccessDeep Offline Reinforcement Learning for Real-world Treatment Optimization Applications
There is increasing interest in data-driven approaches for recommending optimal treatment strategies in many chronic disease management and critical care applications. Reinforcement learning methods are well-suited to this sequential decision-making problem, but must be trained and evaluated exclusively on retrospective medical record datasets as direct online exploration is unsafe and infeasible. Despite this requirement, the vast majority of treatment optimization studies use off-policy RL methods (e.g., Double Deep Q Networks (DDQN) or its variants) that are known to perform poorly in purely offline settings. Recent advances in offline RL, such as Conservative Q-Learning (CQL), offer a suitable alternative. But there remain challenges in adapting these approaches to real-world applications where suboptimal examples dominate the retrospective dataset and strict safety constraints need to be satisfied. In this work, we introduce a practical and theoretically grounded transition sampling approach to address action imbalance during offline RL training. We perform extensive experiments on two real-world tasks for diabetes and sepsis treatment optimization to compare performance of the proposed approach against prominent off-policy and offline RL baselines (DDQN and CQL). Across a range of principled and clinically relevant metrics, we show that our proposed approach enables substantial improvements in expected health outcomes and in consistency with relevant practice and safety guidelines.
- 29 November 2022
- Published by Springer Nature
- p. 29-40
Domain-specific Language Pre-training for Dialogue Comprehension on Clinical Inquiry-Answering Conversations
- 29 November 2022
- Published by Springer Nature
- p. 29-40
Domain-specific Language Pre-training for Dialogue Comprehension on Clinical Inquiry-Answering Conversations
- 7 October 2022
- Published by Springer Nature
- Vol. 13573, 119-129
Towards More Efficient Data Valuation in Healthcare Federated Learning Using Ensembling
Federated Learning (FL) wherein multiple institutions collaboratively train a machine learning model without sharing data is becoming popular. Participating institutions might not contribute equally - some contribute more data, some better quality data or some more diverse data. To fairly rank the contribution of different institutions, Shapley value (SV) has emerged as the method of choice. Exact SV computation is impossibly expensive, especially when there are hundreds of contributors. Existing SV computation techniques use approximations. However, in healthcare where the number of contributing institutions are likely not of a colossal scale, computing exact SVs is still exorbitantly expensive, but not impossible. For such settings, we propose an efficient SV computation technique called SaFE (Shapley Value for Federated Learning using Ensembling). We empirically show that SaFE computes values that are close to exact SVs, and that it performs better than current SV approximations. This is particularly relevant in medical imaging setting where widespread heterogeneity across institutions is rampant and fast accurate data valuation is required to determine the contribution of each participant in multi-institutional collaborative learning. Keywords Federated Learning Data valuation Healthcare AI
- 7 October 2022
- Published by Springer Nature
- Vol. 13573, 119-129
Towards More Efficient Data Valuation in Healthcare Federated Learning Using Ensembling
Federated Learning (FL) wherein multiple institutions collaboratively train a machine learning model without sharing data is becoming popular. Participating institutions might not contribute equally - some contribute more data, some better quality data or some more diverse data. To fairly rank the contribution of different institutions, Shapley value (SV) has emerged as the method of choice. Exact SV computation is impossibly expensive, especially when there are hundreds of contributors. Existing SV computation techniques use approximations. However, in healthcare where the number of contributing institutions are likely not of a colossal scale, computing exact SVs is still exorbitantly expensive, but not impossible. For such settings, we propose an efficient SV computation technique called SaFE (Shapley Value for Federated Learning using Ensembling). We empirically show that SaFE computes values that are close to exact SVs, and that it performs better than current SV approximations. This is particularly relevant in medical imaging setting where widespread heterogeneity across institutions is rampant and fast accurate data valuation is required to determine the contribution of each participant in multi-institutional collaborative learning. Keywords Federated Learning Data valuation Healthcare AI
- 15 September 2022
- Published by Springer Nature
- p. 675-685
Consistency-Based Semi-supervised Evidential Active Learning for Diagnostic Radiograph Classification
- 15 September 2022
- Published by Springer Nature
- p. 675-685
Consistency-Based Semi-supervised Evidential Active Learning for Diagnostic Radiograph Classification