Back to the library

Key question New

AI-processed · Author-approved

Can hospitals predict which patients will seek in-person outpatient care after completing an online medical consultation?

Predicting Online-to-Offline Care Conversion in Internet Hospitals: A Theory-Guided Deep Learning Approach
Yang Zhou, Juan Li, Jiahui Feng, et al. · International Conference on Information Systems (ICIS) · 2026
Illustration for: Can hospitals predict which patients will seek in-person outpatient care after completing an online medical consultation?

Loads Spotify's player; Spotify may set cookies. This device remembers your choice.

Key findings

  • PACC-Net outperformed standard machine learning, neural dialogue architectures, and large language model baselines, achieving an AUC of 0.7738, an AUPRC of 0.7752, and an F1-score of 0.7049.
  • Compared to the strongest baseline (a LoRA fine-tuned Qwen3-4B model), PACC-Net yielded relative improvements of 1.92% in AUC, 3.35% in AUPRC, and 5.05% in F1-score.
  • Ablation analyses confirmed that removing the coping credibility calibrator caused the largest relative reduction in recall (22.44%), while omitting patient vulnerability modeling caused the steepest decline in AUC (1.59%).
  • Sensitivity checks demonstrated consistent performance advantages across 3-day, 7-day, and 14-day follow-up prediction windows.

What it means for you

CIO / IT Executive

Allocate development resources to pilot the PACC-Net framework, prioritizing the integration of multi-turn consultation text, patient historical records, and physician credential data into the core EHR architecture.

IT Manager

Review data pipelines to ensure the team can reliably extract and feed patient threat appraisal variables and physician credibility cues into the prediction models by the end of the sprint.

Business Strategist

Draft a capacity planning proposal that utilizes 3-day, 7-day, and 14-day online-to-offline conversion probability outputs to dynamically adjust physical clinic staffing and appointment slots.

Researcher

Design an ablation experiment to test the impact of integrating Protection Motivation Theory variables against traditional marketing-based features on downstream clinical conversion prediction accuracy.

Policymaker

Draft guidelines for internet hospitals requiring transparent display of physician credentialing data to support patient coping appraisal and improve safe cross-channel care coordination.

Summary

This study develops PACC-Net, a theory-guided deep learning framework that predicts whether a patient will visit an offline clinic within 14 days after an online consultation. Grounded in Protection Motivation Theory, the model evaluates patient threat appraisal and coping appraisal by integrating multi-turn consultation text, prior healthcare utilization records, and physician credentials. The approach was evaluated using a temporal split on linked records covering 65,867 online consultations and 323,196 offline visits from a large tertiary hospital in China.

Problem

Hospitals frequently face imbalances between outpatient capacity and patient demand, yet they lack effective tools to anticipate when online consultations will transition into offline clinic visits. Existing models rely on standard text analysis or marketing frameworks that treat patient inputs uniformly, failing to capture individual differences in health threat perceptions and source credibility cues. Consequently, healthcare providers miss opportunities to proactively coordinate staffing, appointment slots, and cross-channel patient care.

Transcript
Host: Welcome to A.I.S. Insights — powered by Living Knowledge. I'm Anna Ivy Summers. Today we are discussing a study presented at the International Conference on Information Systems 2026, titled "Predicting Online-to-Offline Care Conversion in Internet Hospitals: A Theory-Guided Deep Learning Approach," authored by Yang Zhou, Juan Li, Jiahui Feng, and Hefu Liu. Alex, this study examines what happens after a patient completes an online consultation.

Expert: Thanks, Anna. In many healthcare systems, online consultations do not simply replace in-person visits. Instead, they often serve as an initial entry point into a broader care pathway, where a subset of patients subsequently visits an outpatient clinic. The central objective of this study is predicting whether a patient will seek offline care within 14 days following an online consultation.

Host: Why is predicting this transition so important for hospital management?

Expert: Hospitals consistently face imbalances between outpatient capacity and patient demand. If administrators can anticipate which online consultations will convert into offline clinic visits over the subsequent two weeks, they can aggregate these forecasts by department. This gives them an early indicator of demand to inform clinical staffing, adjust appointment slot allocations, and coordinate follow-up care across channels.

Host: Prior analytical models have struggled with this task. Why didn't standard language models or general prediction methods suffice?

Expert: Most earlier conversion models were designed for marketing contexts, where actions reflect immediate consumer preference. In healthcare, an offline visit is a protective health behavior. Standard text models typically process dialogue uniformly across all individuals, ignoring prior utilization history and variations in how patients interpret their symptoms. They also generally fail to account for how physician credibility influences whether a patient follows medical advice.

Host: To address that, the authors grounded their design in Protection Motivation Theory. How did they translate that theory into a neural network?

Expert: Protection Motivation Theory posits that protective behaviors stem from two cognitive processes: threat appraisal and coping appraisal. The authors developed PACC-Net, which builds two dedicated branches to reflect these processes. For threat appraisal, the model combines the consultation dialogue with the patient's prior hospital visit history to estimate vulnerability, then uses a mixture-of-experts module to evaluate perceived symptom severity. For coping appraisal, it extracts recommended coping cues from the dialogue and dynamically calibrates them with physician credentials, including the doctor's title, department, and background.

Host: Tell us about the empirical evaluation and the dataset they used.

Expert: The authors evaluated the model using linked records from a large tertiary public hospital in Hefei, China, covering January through June 2024. The dataset contained 65,867 completed online consultations and 323,196 offline outpatient records. To reflect operational deployment and avoid data leakage, they used a temporal split: training on data from January to May and evaluating on June.

Host: What were the key findings from their benchmark tests?

Expert: PACC-Net consistently outperformed traditional machine learning models, dialogue architectures, and large language model baselines. It achieved an AUC of 0.7738, an AUPRC of 0.7752, and an F1-score of 0.7049. Relative to the strongest baseline, which was a LoRA fine-tuned Qwen3-4B model, PACC-Net showed relative improvements of 1.92% in AUC, 3.35% in AUPRC, and 5.05% in F1-score.

Host: What did the ablation analyses reveal about the individual model components?

Expert: The ablation tests showed that removing the Coping Credibility Calibrator caused the steepest decline in recall, dropping it by 22.44%. Removing the Patient Vulnerability Modeling caused the largest decline in AUC, at 1.59%. Sensitivity analyses also showed that PACC-Net maintained its predictive edge across 3-day, 7-day, and 14-day observation windows.

Host: Looking at the business and operational impact, what should hospital leaders take away from this?

Expert: The primary operational takeaway is that hospitals can utilize data already present in their systems—dialogue logs, visit histories, and physician profiles—to generate actionable short-term demand signals. By anticipating offline conversion at the department level, operations teams can shift from reactive scheduling to proactive resource planning, optimizing clinic capacity while supporting coordinated patient care.

Host: That provides a clear perspective on connecting online consultations to physical operations. Alex, thank you for sharing these insights, and thank you to our listeners for tuning in to A.I.S. Insights — powered by Living Knowledge.
Artificial Intelligence & Automation Information Systems & Management Human-Computer Interaction Data Analytics & Big Data Digital Transformation & Strategy Cybersecurity & Privacy Cloud & Infrastructure E-Commerce & Digital Platforms Operations & Supply Chain Sustainability & Green IS

Online-to-Offline Care Conversion, Internet Hospitals, Protection Motivation Theory, Deep Learning, Outpatient Demand Prediction, Healthcare Analytics