AI-Based Precision Prediction of Healthcare Workers’ Antibiotic Use Intentions: Multi-theoretical Psychological and Behavioral Modeling

Le Han, Peng Xian, Xiao Liu, Ying Liu, Yue Chang, Lijun Ma, Jie Ren, Yunshuang Liu, Xiaosong Wang, Yujuan Long, Yuan Zhou, Li Ren, Fang Jin, Ling Liang, Hui Tang, Qi LiuResearch Square 2025 · DOI

Abstract

Introduction: Antimicrobial resistance (AMR) remains a global health priority, partly driven by irrational antibiotic prescribing among healthcare workers. Existing interventions often overlook the psychological and contextual factors that shape prescribing behaviour. This study aimed to integrate behavioural theory with explainable machine learning to identify psychological predictors of antibiotic use intention among clinicians. Methods: A cross-sectional survey was conducted among 1,135 healthcare workers from four public hospitals in China. Participants completed questionnaires based on constructs from the Theory of Planned Behavior, Health Belief Model, Rational Action Theory, Self-Efficacy Theory, Social Support Theory, and cognitive processing frameworks. LASSO regression and SHAP analysis were applied alongside machine learning classifiers (e.g. XGBoost, LightGBM, CatBoost) to identify key predictors and interactions influencing prescribing intention. Results: Social support, cognitive processing, knowledge and skills, and health beliefs were the most important predictors. SHAP analysis revealed nonlinear interactions, particularly between social support and cognitive engagement. Ensemble models achieved high predictive accuracy (F1-scores >0.94) for high and medium prescribing intention, but classification of low-intention individuals remained more challenging. Conclusion: Combining behavioural theory with explainable AI offers a scalable approach to identifying clinicians at risk of irrational prescribing. These findings support the development of psychologically tailored, real-time interventions that can improve antibiotic stewardship and address AMR in diverse health system settings.

Introduction

Antimicrobial resistance (AMR) is often described as a "silent pandemic," largely fueled by the irrational use of antibiotics. While clinical guidelines and audit systems are standard, they frequently fail because they treat prescribing as a purely logical, knowledge-based task. In reality, a clinician’s decision to prescribe is a complex human process influenced by diagnostic uncertainty, time pressure, and perceived patient expectations.

Existing research into these behaviors has traditionally been limited by narrow theoretical scopes and linear statistical methods that fail to capture complex, real-world interactions. This study bridges that gap by integrating six major psychological theories with explainable machine learning to predict healthcare workers' intentions to prescribe antibiotics rationally.

Methods

The researchers developed a "theory-meets-data" pipeline consisting of:

  1. Multi-theoretical Modeling: A 32-item scale based on eight latent constructs from theories including the Theory of Planned Behavior (TPB), Health Belief Model (HBM), and Social Support Theory.
  2. Feature Selection: LASSO regression was used to identify the most impactful psychological predictors.
  3. Machine Learning Classifiers: Eleven algorithms were trained, including XGBoost, LightGBM, and CatBoost, alongside ensemble methods like Stacking and Voting.
  4. Explainability: SHAP (SHapley Additive exPlanations) values were used to visualize how psychological factors contributed to predictions and their interactions.

Study Population

The study utilized a primary dataset from a cross-sectional survey of 1,135 healthcare workers across four public hospitals in China between January and December 2024.

Demographic characteristics of the study participants (N=1,135).

As shown in the table above, the cohort was predominantly female (81.32%), with approximately 66% aged between 26 and 45. Participants represented a diverse range of departments, including internal medicine, surgery, and intensive care.

Key Insights and Model Performance

The analysis revealed several critical findings regarding the drivers of prescribing behavior:

  • Context Over Willpower: Traditional "volitional" constructs like self-efficacy showed weak associations with rational prescribing. In regulated hospital environments, institutional norms often override individual agency.
  • The Power of Social Support: Social support (institutional culture and peer reinforcement) was the strongest predictor (beta = 0.2564).
  • The "Amplifier" Effect: SHAP analysis showed that the positive effect of social support is significantly amplified in individuals with high cognitive engagement. Essentially, a supportive environment works best for doctors who are already actively "thinking through" their prescribing decisions.
  • Predictive Accuracy: Ensemble models achieved F1-scores > 0.94 and AUC values of 0.99 for "High" and "Medium" prescribing intentions. However, predicting "Low" rational prescribing intention was more challenging (F1-scores of 0.76–0.80), likely due to external stressors like workload.

Discussion and Future Directions

While the study identifies strong associations, it relies on self-reported intentions rather than actual prescribing data. Furthermore, the cross-sectional design and geographic focus on China suggest a need for broader validation in different cultural and economic contexts.

The ultimate goal of this research is to develop an "intelligent system" capable of delivering tailored, real-time psychological interventions to clinicians. By moving from generic guidelines to precision behavioral health, antimicrobial stewardship can more effectively address the human factors driving AMR.

Summary updated

Abstract

<title>Abstract</title> Introduction: Antimicrobial resistance (AMR) remains a global health priority, partly driven by irrational antibiotic prescribing among healthcare workers. Existing interventions often overlook the psychological and contextual factors that shape prescribing behaviour. This study aimed to integrate behavioural theory with explainable machine learning to identify psychological predictors of antibiotic use intention among clinicians. Methods: A cross-sectional survey was conducted among 1,135 healthcare workers from four public hospitals in China. Participants completed questionnaires based on constructs from the Theory of Planned Behavior, Health Belief Model, Rational Action Theory, Self-Efficacy Theory, Social Support Theory, and cognitive processing frameworks. LASSO regression and SHAP analysis were applied alongside machine learning classifiers (e.g. XGBoost, LightGBM, CatBoost) to identify key predictors and interactions influencing prescribing intention. Results: Social support, cognitive processing, knowledge and skills, and health beliefs were the most important predictors. SHAP analysis revealed nonlinear interactions, particularly between social support and cognitive engagement. Ensemble models achieved high predictive accuracy (F1-scores &gt;0.94) for high and medium prescribing intention, but classification of low-intention individuals remained more challenging. Conclusion: Combining behavioural theory with explainable AI offers a scalable approach to identifying clinicians at risk of irrational prescribing. These findings support the development of psychologically tailored, real-time interventions that can improve antibiotic stewardship and address AMR in diverse health system settings.
BibTeX
@inproceedings{han2025ai,
  author = {Le Han and Peng Xian and Xiao Liu and Ying Liu and Yue Chang and Lijun Ma and Jie Ren and Yunshuang Liu and Xiaosong Wang and Yujuan Long and Yuan Zhou and Li Ren and Fang Jin and Ling Liang and Hui Tang and Qi Liu},
  title = {{AI-Based Precision Prediction of Healthcare Workers’ Antibiotic Use Intentions: Multi-theoretical Psychological and Behavioral Modeling}},
  booktitle = {Research Square},
  year = {2025},
  doi = {10.21203/rs.3.rs-6768023/v1},
}
Figure