Application, challenges and prospects of artificial intelligence in acute poisoning management

Abstract

BACKGROUND Acute poisoning remains a major global public health challenge; its management is often limited by diagnostic uncertainty, narrow therapeutic windows, mixed or unknown exposures, and uneven access to toxicology expertise. Although the use of artificial intelligence (AI) in medicine has been increasingly explored, its role in acute poisoning management warrants a focused review.

METHODS We conducted a narrative review of publications indexed in PubMed, Web of Science, and Wanfang Data from January 1, 2021 to March 1, 2026. Eligible publications addressed AI applications in acute poisoning. The included studies were driven by machine learning (ML) and deep learning (DL) architectures (e.g., artificial neural networks [ANNs]), employed technical modalities such as natural language processing (NLP), computer vision (CV), or large language models (LLMs), and utilized clinical decision support systems (CDSSs) or geographic information system (GIS)-assisted tools. Studies that incorporated explainable AI (XAI) methodologies were also included.

RESULTS Current evidence indicates that AI has been used for medical history acquisition, toxin identification, individualized treatment support, complication warning, prognostic assessments, remote consultations, emergency responses, and public education. Most tools, however, remain at the model development, retrospective validation, laboratory testing, or small-scale feasibility evaluation stage. Representative quantitative findings include the identification of CV-based snakes with a species-level accuracy of 96.0% and a genus-level accuracy of 99.0%; additionally, in selected poisoning scenarios, extreme gradient boosting (XGBoost)-based prognostic models have yielded area under the curve (AUC) values above 0.907. On the basis of this review, a three-stage framework and a full-process perspective are proposed for clinically grounded AI development.

CONCLUSION AI has substantial potential across the acute poisoning care continuum. Future translation will require multicenter data, external validation, interpretability, workflow integration, and ethical and regulatory oversight.

INTRODUCTION

Acute poisoning remains a major global public health challenge and a leading cause of preventable morbidity and mortality; it is characterized by acute onset, complex toxicological mechanisms, and rapid clinical progression.^[1,2] Early laboratory testing and imaging often lack sufficient diagnostic specificity, and many patients cannot provide a reliable history of toxicity exposure. These factors contribute to missed diagnoses, delayed diagnoses, and misdiagnoses.^[3] Management is complicated by the use of novel toxicants, mixed exposures, limited antidote availability, and an uneven diagnostic and treatment capacity across medical settings.^[4,5]

Conventional diagnostic and treatment strategies are often limited by delayed responses, time-consuming data collection, and reliance on individual clinical experience.^[6] Artificial intelligence (AI) can be used to extract clinically relevant information from electronic health records (EHRs), assist in toxicant identification, support risk stratification, and provide individualized decision support.^[7-10] In acute poisoning, these functions are particularly relevant because management decisions are often made with incomplete exposure histories and rapidly changing clinical conditions.

Previous studies^[5,11] have summarized AI applications in emergency toxicology and broader toxicological sciences. However, few studies have constructed a three-step developmental framework for AI in acute poisoning management or have proposed a full-process paradigm spanning prehospital emergency care, in-hospital diagnosis and treatment, and public health response. This review fills this gap by integrating clinical scenarios with technological maturity and translational needs. We summarize recent advances in AI applications across the full spectrum of acute poisoning management, including differential diagnosis, treatment support, prognostic assessment, emergency response, and public health intervention. The objectives of this review are to propose a three-stage framework for AI development; synthesize evidence across the prehospital, in-hospital, and public health continuum; and outline a clinically oriented pathway for future translation.

METHODS

Study design

This study was designed as a narrative review. The aim of this study was to provide a clinically oriented synthesis of recent AI applications and future directions for acute poisoning management.

Data sources and search strategy

A comprehensive literature search was conducted in PubMed, Web of Science, and Wanfang Data. The inclusion criteria were as follows: (1) studies focused on AI technologies in acute poisoning or closely related emergency toxicology scenarios; (2) studies addressing toxicant identification, medical history acquisition, treatment support, complication warning, prognostic assessments, remote care, emergency responses, or public education; and (3) original studies, systematic or narrative reviews, clinical guidelines, expert consensus statements, or high-quality conference papers.

The exclusion criteria were as follows: (1) case reports, letters, editorials, comments, and abstracts without sufficient methodological information; (2) studies unrelated to poisoning or clinical toxicology; and (3) studies not involving AI-based methods.

The complete Boolean search formula was as follows: (“artificial intelligence” OR “machine learning” OR “deep learning” OR “natural language processing” OR “computer vision” OR “large language model” OR “clinical decision support”) AND (“acute poisoning” OR “clinical toxicology” OR “emergency toxicology” OR “overdose” OR “intoxication” OR “toxic agent identification” OR “poison management”) AND (“diagnosis” OR “prognosis” OR “treatment” OR “triage” OR “risk prediction” OR “telemedicine” OR “remote care” OR “public health”).

The primary search period was January 1, 2021, to March 1, 2026, to capture recent advances in AI applications for acute poisoning management. For topics with limited direct evidence, such as poisoning-specific clinical decision support systems, treatment-support tools, and foundational AI-assisted toxicology methods, the search was extended to 2017 only when studies were directly relevant, methodologically important, or frequently cited in subsequent literature. This extension was applied according to predefined relevance criteria rather than ad hoc selection.

Chinese-language studies retrieved from Wanfang Data were screened using the same eligibility criteria as English-language studies. Key information, including study design, population, AI model type, poisoning category, performance metrics, and conclusions, was extracted and synthesized with English-language evidence. Chinese-language studies were used mainly to supplement evidence from regional clinical practice and toxicology-specific applications. A total of 3,251 references were identified in the preliminary search. These were subsequently refined to 45 publications based on predefined criteria, including impact factor and academic importance. Furthermore, four articles pertinent to public health were manually supplemented, culminating in a final collection of 49 articles.

The methodological quality of the included literature was assessed using the Joanna Briggs Institute (JBI) critical appraisal checklists when applicable. A total of 49 publications were included in the final qualitative synthesis.

APPLICATION OF AI IN THE PREHOSPITAL MANAGEMENT OF ACUTE POISONING

AI-driven chatbots combined with geographic information systems (GISs) may improve poisoning-related emergency responses. Integrated systems using geolocation and ambulance tracking can help users locate nearby emergency services and initiate emergency calls, thereby shortening prehospital response times.

Natural language processing (NLP)-based chatbots can interpret user-reported symptoms and emergency scenarios and provide context-aware first-aid guidance.^[12,13] These tools may help bystanders provide standardized first aid before professional responders arrive, but they should include escalation prompts for severe or uncertain presentations.

In large-scale poisoning incidents, rapid emergency response and toxicity dispersion prediction are essential for reducing casualties. ML models can estimate the spatial impact of toxic dispersion using toxicant properties and environmental conditions.^[14] These predictions may support disaster management authorities in response planning and medical resource allocation.

APPLICATIONS OF AI IN THE IN-HOSPITAL MANAGEMENT OF ACUTE POISONING

Applications of AI in medical history acquisition and toxic agent identification

The early and accurate identification of toxicants is important for acute poisoning management because it determines the timing of intervention and subsequent treatment. In emergency settings, clinicians often make etiological judgments with fragmented histories and limited time. AI-based tools may help reduce information loss, support rapid triage, and take part in identifying likely toxicants within the therapeutic window.

Intelligent acquisition of medical history data via NLP

In emergency settings, patients may have impaired consciousness, and information provided by family members or bystanders may be incomplete. These factors can result in fragmented, unstructured histories and delayed diagnosis. NLP can be used to extract toxicant names, exposure routes, symptom trajectories, and medication histories from EHRs, poison-center calls, family narratives, and nursing documentation.

Recent toxicology AI literature emphasizes the feasibility of converting heterogeneous toxicological information into structured computational inputs.^[14-16] In the acute poisoning workflow, this capability is most relevant to exposure-history collection, poison-center consultation records, and EHR narratives, where NLP can be used to extract toxicant names, exposure routes, timing, symptoms, and coexposures for downstream analysis. Compared with structured-data models, NLP is most useful when exposure information is dispersed across free-text records, but its performance may be reduced by colloquial descriptions, missing temporal details, mixed exposures, and institution-specific documentation styles. Therefore, NLP outputs should be checked against bedside history, laboratory findings, and toxicology screening before they are used for clinical inference.

Precise toxic agent identification using machine learning (ML) and deep learning (DL)

The diversity of toxicants and the variable relationship between exposure and symptoms make etiological inference difficult. ML and DL models can be used to identify discriminative patterns in high-dimensional clinical data and may support toxicant classification. Explainable artificial intelligence (XAI) methods may improve transparency by revealing which variables contribute to a prediction.

With respect to toxicant classification, several studies have reported better performance for ML-based or DL-based systems than for conventional rule-based approaches in selected settings, but the evidence should be interpreted according to case complexity and validation design. In 300 simulated poisoning cases, the Tox diagnostic system achieved Cohen’s kappa values of 0.8432, 0.4396, and 0.3331 for simple, moderately complex, and highly complex cases, respectively.^[17] ToxNet, which was trained on 781,278 single-exposure cases from U.S. poison control centers, achieved an F1 score of 0.89±0.02 for toxicant classification, compared with a clinical baseline of 0.72±0.05.^[18]

For common toxicant categories, ML models have also shown high specificity. A classification algorithm trained on 201,031 cases from the U.S. National Poison Data System and targeting eight common toxicant categories achieved an overall specificity of 92%; the specificity exceeded 99% for selected agents, including sulfonylureas and calcium channel blockers.^[19] In resource-limited settings, an ensemble learning-based pesticide poisoning diagnostic system integrated symptoms and exposure history, yielding an accuracy of 96% for preliminary screening.^[20,21] These results suggest potential value for early triage, but prospective validation in real emergency workflows is still needed.

Clinically, AI toxicant identification models are adjuncts, not replacements. Simple models are transparent but limited by predefined variables, while deep learning captures complex patterns but requires large datasets and suffers from poor generalizability. Although explainable AI (XAI) enhances trust, it cannot eliminate data bias. Performance varies significantly across settings because of differences in data quality, case complexity, and validation methods. Thus, AI must complement—not substitute—clinical judgment grounded in laboratory results, toxicology screens, and exposure history.

Toxic agent identification via computer vision (CV)

In onsite emergency settings, visual information may include drug packaging, chemical container labels, venomous animals, bite wounds, and characteristic clinical signs. Conventional manual assessments can be subjective and may delay decision-making.

CV enables standardized analyses of visual information and may support the rapid identification of toxins in prehospital and onsite settings. With respect to snakebite management, a model trained on 386,066 images achieved a species-level accuracy of 96.0% and a genus-level accuracy of 99.0%; thus, this model may help clinicians select the appropriate antivenom when expert identification is unavailable.^[22]

Applications of AI in acute poisoning treatment

The goal of acute poisoning treatment is precise, efficient, and individualized intervention. However, the toxicant dose, exposure route, comorbidities, age, organ function, and time from exposure to treatment may affect the treatment response. AI-assisted strategies have therefore been explored to optimize antidote use, integrate clinical decision support systems (CDSSs), and predict complications.

Individualized optimization of antidote dosage and timing

Precise antidote administration is central to acute poisoning treatment. Both underdosing and overdosing may reduce therapeutic efficacy or cause iatrogenic harm. AI-based systems may integrate patient characteristics and exposure-related data to optimize the dose, duration, and timing of antidote administration, but such recommendations should be reviewed by clinicians before implementation.

With respect to opioid poisoning, an extreme gradient boosting (XGBoost)-based model predicted the maintenance dose and duration of naloxone administration, with an area under the receiver operating characteristic curve (AUC) of 0.97.^[23] More broadly, adaptive learning and predictive modeling may inform future antidote-dosing strategies, but poisoning-specific models require prospective safety evaluations before clinical use.^[24]

At the system level, an antidote application system that integrates 27 algorithms covering 29 common antidotes has been developed as an educational tool. By incorporating parameters such as body weight, renal function, and exposure level, such systems may translate toxicological knowledge into actionable dosing guidance and may reduce medication errors when used under clinical supervision.^[25] In critical care, ML models have also been explored for serum drug concentration prediction and medication dosing support using vital signs, laboratory results, and pharmacokinetic parameters.^[26]

Intelligent integration of CDSSs

In complex poisoning cases, AI-driven CDSSs can be used to integrate diagnosis, treatment recommendations, early warning, and knowledge retrieval. These systems can be used to analyze multisource clinical data and provide individualized decision support, but their outputs should be reviewed by clinicians before being used in time-sensitive decisions. To address black-box limitations and improve clinician trust, CDSS architectures increasingly incorporate XAI tools, including SHapley Additive exPlanations (SHAP) and local interpretable model-agnostic explanations (LIME), to display the clinical variables that contribute to AI-generated recommendations.^[27] Such architectures can support clinician feedback loops. A clinician-in-the loop design allows physicians to verify, adjust, or reject AI-generated recommendations, thereby preserving clinical responsibility and reducing the risk of automation bias. In emergency settings, poisoning-specific CDSSs have been developed to generate treatment recommendations by integrating and processing multisource data.^[28]

Recent studies have explored the integration of large language models (LLMs) with real-world emergency data. By synchronizing EHRs, bedside monitoring data, and laboratory results, these systems can construct patient profiles, support risk stratification, and generate preliminary treatment suggestions.^[29] However, LLM-based recommendations should be linked to curated toxicology knowledge and reviewed by clinicians because factual errors, hallucinations, and outdated information remain important safety concerns. LLM-based toxicology support should be used with particular caution in emergency settings. These systems perform poorly for rare or newly emerging toxicants and provide inconsistent answers when the same scenario is phrased differently. Practical safeguards should include retrieval-augmented generation from validated toxicology guidelines, source attribution, uncertainty warning, escalation prompts, clinician confirmation, and audit logging.

AI-driven early warning of complications and clinical deterioration for risk-stratified decision-making

Acute kidney injury, respiratory failure, and cardiac conduction abnormalities are major complications of acute poisoning. Early identification and timely intervention are important for improving outcomes. AI models that use real-time physiological data and historical clinical cases may support early warning of complications and critical illness progression.

In acute pesticide poisoning, a long short-term memory (LSTM) model predicted acute respiratory failure with a positive predictive value of 98.42% and a sensitivity of 97.91%.^[30] In methanol poisoning, an ensemble learning model based on 897 clinical cases predicted the need for tracheal intubation with a specificity of 99% and a sensitivity of 100%.^[31] In terms of atypical antipsychotic poisoning, eight ML models were used to predict QTc prolongation and the need for mechanical ventilation in 355 patients; as a predictor of a prolonged QTc interval, all three-based models achieved perfect specificity, recall, precision, accuracy, and an AUC of 100% using the training dataset.^[32] These findings suggest potential value for early intervention, although calibration and decision curve analyses should be reported more consistently.

In clinical use, treatment-oriented AI systems have different strengths. Dose-prediction models such as XGBoost are useful for narrowly defined decisions, whereas CDSSs can integrate guidelines, laboratory data, and clinician feedback across a broader workflow. LLM-based systems are more flexible for knowledge retrieval and summarization, but they carry greater risks of hallucination, outdated recommendations, and inconsistent responses.^[23,28-32] Differences in reported performance across dosing, CDSS, and complication warning studies should therefore be interpreted in relation to the endpoint, data availability, patient severity, and validation design. Antidote dosing also has narrow safety margins; AI recommendations should be confirmed by clinicians and interpreted cautiously for children, pregnant patients, patients with hepatic or renal dysfunction, and patients with mixed-exposure cases. These tools should be embedded in a CDSS rather than used to generate automatic medical orders.

Applications of AI in the prognostic evaluation of acute poisoning

Acute poisoning may progress rapidly and shows marked interindividual heterogeneity. These features make disease progression difficult to predict. AI models based on multidimensional clinical data may not only support early complication warning but also allow for mortality risk stratification and prognostic assessment, thereby helping clinicians prioritize intervention and allocate resources.

Stratified assessment of overall prognosis and mortality risk

Accurate prognosis and mortality risk stratification can help prioritize medical resources for high-risk patients. However, risk alerts should be clinically validated and interpreted together with bedside assessments.^[33]

Several studies have reported that AI models may outperform conventional statistical approaches in selected prognostic tasks. An artificial neural network (ANN) model based on data from 1,304 patients had a higher AUC for in-hospital mortality than logistic regression did.^[34] A gradient boosting classifier achieved an AUC of 0.947 for clinical outcome stratification, with the methanol dose and Glasgow Coma Scale score identified as important predictors.^[35] For specific poisoning types, XGBoost and categorical boosting (CatBoost) models have been applied to calcium channel blocker poisoning,^[36] and a decision tree-based model has provided interpretable clinical pathways for the prognostic assessment of acute paraquat poisoning.^[37]

Evaluation of severity and delayed sequelae

Some toxicant exposures are associated with delayed progression and long-term sequelae. Predictive assessments may support early intervention and long-term follow-up planning. With respect to the severity assessment of acute organophosphate poisoning, the XGBoost model achieved an AUC of 0.907.^[38] For delayed neurological sequelae after acute carbon monoxide poisoning, an optimized synthetic minority oversampling technique-random forest (SMOTERF) model achieved an AUC of 0.89.^[39] Related studies have also shown the potential of random forest models, and SHAP analysis may improve interpretability by identifying important predictive features for delayed encephalopathy after carbon monoxide poisoning.^[40]

Prognostic models for specific high-risk toxicants and clinical perspective

AI-based prognostic models have been explored for several high-risk toxicants associated with rapid deterioration or severe organ injury. Logistic regression models supported by AI-assisted analysis have shown good calibration for predicting a prolonged ICU stay for patients with drug poisoning.^[41] In acute diquat poisoning, a support vector machine (SVM) model improved mortality prediction.^[42] In terms of methadone poisoning, the Light Gradient Boosting Machine (LightGBM) and XGBoost models achieved high specificity and precision in terms of outcome prediction.^[43] These findings suggest potential value for selected high-risk poisoning scenarios, but broader external validation is needed before routine clinical use.

From a clinical perspective, prognostic models may help identify high-risk patients, guide intensive care unit (ICU) triage, and support resource allocation. Model choice should reflect the clinical question. Logistic regression and decision-tree models are easier to interpret, whereas ANN, gradient boosting, and XGBoost models may better capture nonlinear interactions among dose, time to treatment, vital signs, electrocardiographic changes, and laboratory abnormalities.^[34-41]

XAI methods may bridge this gap by providing feature-attribution information, but they do not replace external validation. Apparent inconsistencies across reported AUC values and model rankings probably reflect heterogeneity in toxicant type, outcome definition, class imbalance, sample size, and validation strategy. Therefore, prognostic studies should report calibration, decision curve analysis, subgroup performance, and external validation results rather than discrimination results alone.

APPLICATIONS OF AI IN REMOTE EMERGENCY CARE AND PUBLIC HEALTH INTERVENTIONS FOR ACUTE POISONING

The integration of AI with telemedicine, GIS, and LLMs may reduce barriers related to geographic distance and uneven medical resources. These approaches can support remote emergency care, public health response, and poisoning prevention education. Accordingly, AI-assisted poisoning management extends from hospital-based care to out-of-hospital and population-level scenarios.

Remote emergency consultation and real-time guidance

Geographic barriers remain important obstacles to timely treatment for patients with acute poisoning in remote areas. AI-enabled telemedicine may support the cross-regional sharing of specialist resources and improve emergency care capacity in underresourced settings.

For example, a localized tele-ICU model has been developed in India for severe envenomation, including snakebites. Through vital sign monitoring, cloud-based data synchronization, and video consultation, toxicology specialists can guide frontline clinicians in standardized treatment.^[44] LLMs have also been explored as knowledge-support tools for primary care, although pediatric poisoning and complex mixed exposures still require domain-specific optimization and specialist oversight.^[45]

Intelligent public education for poisoning prevention

Reducing the incidence of poisoning at the source is a central goal of prevention. LLM-based systems may generate tailored education materials for households, schools, rural communities, and workplaces and may improve the accessibility of poisoning prevention information.

Studies comparing AI-generated answers with toxicologist responses suggest that LLM-based systems may produce readable and clinically relevant explanations for poisoning symptoms, treatment principles, and prevention measures.^[46,47] Public-facing tools, however, must avoid unsupported treatment advice and should direct users to emergency medical services or poison centers when risk is unclear.

In real-world implementation, remote consultation, GIS-assisted response, chatbot-based first-aid guidance, and LLM-based public education serve different purposes. Telemedicine is most relevant for connecting remote clinicians with toxicology expertise; GIS tools support dispatch and population-level emergency planning, and LLMs may improve standardized information delivery.^{[12-14,44-47]} Their results are not directly comparable because the reported endpoints differ substantially, including consultation feasibility, response time, public readability, simulated guidance quality, and clinical outcome prediction. This heterogeneity explains why apparently favorable performance in one setting should not be generalized to another. These tools should therefore be used for triage support, standardized communication, and access improvement but not as replacements for emergency medical services, poison centers, or bedside clinical judgment.

Representative studies, including clinical scenarios, poisoning categories, AI methods, sample sizes, model performance, validation types, and major clinical limitations, are summarized in supplementary Table 1. These metrics should be interpreted cautiously because many studies used retrospective, simulated, single-center, or category-specific datasets and lacked multicenter external validation.

CHALLENGES

Despite these advances, there are still three interrelated challenges regarding clinical translation. First, data-related limitations remain central. Many studies are based on retrospective or single-center datasets, with incomplete exposure histories, heterogeneous laboratory measurements, and limited representation of mixed poisoning, rare toxicants, pediatric patients, and resource-limited settings. Multicenter poisoning registries, common data elements, and standardized outcome definitions are therefore needed to support external validation and continuous model updating. Second, AI models must be safe, interpretable, and robust under emergency uncertainty. In acute poisoning, delayed presentation, missing data, novel psychoactive substances, and unfamiliar exposure patterns may reduce model reliability. Therefore, model calibration, uncertainty estimation, out-of-distribution detection, and XAI should be reported alongside discrimination metrics. LLM-based tools should also be constrained by validated toxicology resources and should provide clear escalation prompts when specialist consultation is needed. Third, clinical implementation should follow a staged pathway rather than immediate deployment. Candidate systems should progress from retrospective validation to multicenter external validation, silent prospective testing, pilot implementation, and pragmatic clinical trials. Clinician-in-the-loop designs, accountability, audit trials, automation-bias mitigation, and postdeployment monitoring are needed before routine use.^[48,49] Table 1 provides a concise overview of these barriers and implementation priorities.

Table 1.

Key challenges and implementation priorities for AI translation in acute poisoning management

Dimension	Main barrier	Priority action	Implementation focus
Data foundation	Retrospective, single-center, incomplete, or non-standardized poisoning data	Build multicenter registries and common data elements	External validation, subgroup analysis, continuous updating
Model safety	Limited interpretability, calibration, robustness, and uncertainty reporting	Use explainable AI, calibration, uncertainty estimation, and out-of-distribution detection	Clinician-readable outputs and safer decision support
Clinical implementation	Workflow mismatch, automation bias, unclear accountability, and limited regulatory evidence	Adopt clinician-in-the-loop design and staged validation	Silent prospective testing, pilot deployment, pragmatic trials, audit trails

AI: artificial intelligence.

CONCLUSION

AI applications in acute poisoning management have begun to address several unmet clinical needs, including early differential diagnosis, individualized treatment support, CDSSs, prognostic assessments, remote emergency care, and public health interventions. As seen in Figure 1, these applications extend AI-assisted poisoning management beyond the hospital by connecting prehospital response, remote consultation, emergency dispatch, and public health education. In Figure 2, most remote and public health tools currently remain between stage 1 and stage 2; they may approach stage 3 only when they are linked to validated toxicology knowledge bases, real-time geospatial data, clinician supervision, and postdeployment monitoring. Future development should remain clinically grounded and should prioritize tools that improve diagnostic timeliness, risk stratification, workflow efficiency, and safe decision support across the poisoning care continuum.

Figure 1. Full-process framework of AI-assisted acute poisoning management. The figure summarizes AI applications across the pre-hospital, in-hospital, and public-health continuum, including emergency response, toxicant identification, clinical decision support, prognostic assessment, remote consultation, emergency dispatch, and public education. AI: artificial intelligence; NLP: natural language processing; ML: machine learning; DL: deep learning; XAI: explainable artificial intelligence; ICU: intensive care unit; LLM: large language model; GIS: geographic information system; XGBoost: extreme gradient boosting; CatBoost: categorical boosting; SVM: support vector machine; SHAP: SHapley Additive exPlanations; CDSS: clinical decision support systems.

Figure 2. Three-stage developmental framework for AI in acute poisoning management. Stage 1 represents basic assistance and information structuring; stage 2 represents integrated prediction and decision support; and stage 3 represents clinically embedded, prospectively validated, and continuously monitored intelligent decision systems. AI: artificial intelligence; ICU: intensive care unit.

Funding: This research received no funding.

Ethical approval: Not needed.

Conflicts of interest: All authors declared no conflicts of interest related to this manuscript.

Contributors: SWH: literature search, framework design, draft writing, and proposal of the three-stage AI application framework; XRC: literature search, introduction optimization, terminology standardization, and figure revision; ZQL: academic guidance, project management, manuscript revision, and content validation.

All the supplementary files in this paper are available at http://wjem.com.cn.