Systematic Review of the Intraoperative Hypothermia Risk Prediction Models in Total Joint Arthroplasty Patients

Introduction

Intraoperative hypothermia (IH) is a common surgical complication, defined as a core body temperature below 36°C during surgery.¹ It is one of the most prevalent adverse events, and the prevention of perioperative hypothermia is crucial for improving surgical safety and nursing quality.² Methods for maintaining patient temperature during surgery include forced-air warming, fluid warming, and thermal blankets.³ The incidence of IH in TJA ranges from 10% to 72.6%.⁴ One study in the United States found that 72.6% of TJA patients experienced hypothermia, with 20.6% of them having prolonged hypothermia lasting more than one hour.⁵ A study from the United Kingdom reported a higher 30-day mortality in hip arthroplasty patients with perioperative hypothermia, suggesting that hypothermia increases the risk of death.⁶ Preventing and managing hypothermia is a critical aspect of improving perioperative nursing safety. However, there are still significant differences in preventive and clinical assessment methods for intraoperative hypothermia, and emergency protocols for intraoperative hypothermia are often lacking.⁷ Early identification of high-risk factors for intraoperative hypothermia is a key task in modern operating room nursing.

Current clinical tools for hypothermia—including the Predictors score,⁸ Swiss Staging Model, and Cold Discomfort Scale (CDS)—face significant limitations. These scales often exhibit low predictive power (eg., Predictors score AUC 0.789), overestimation of risk, or insufficient sensitivity to dynamic intraoperative thermal changes.^9–12 While Machine Learning (ML) has been integrated into clinical risk prediction to address these gaps,13-21 existing models still show substantial heterogeneity in data sources and algorithms.

This research systematically reviews TJA-related hypothermia prediction models. By assessing performance and quality, it offers evidence-based guidance for perioperative clinical implementation.

Material and Methods

Search Strategy

A systematic search was conducted in PubMed, Embase, Cochrane Library, Web of Science, CINAHL, Wanfang, China National Knowledge Infrastructure (CNKI), VIP databases, and SinoMed, and relevant references were traced to supplement the search. The search period covered from database inception to August 30, 2025. The search strategy for the English databases, exemplified by PubMed (Figure 1).

Figure 1 PubMed search strategy. The symbol “#” designates a search set identifier, used to track discrete search operations within the database retrieval process. The asterisk “*” serves as a truncation/wildcard operator, representing any number of characters (including zero). It is applied to capture various inflections and derivatives of the same root word, ensuring an extensive and comprehensive literature search.

Inclusion and Exclusion Criteria

Inclusion Criteria

① Study participants were patients scheduled for total hip/knee arthroplasty, aged ≥18 years. ② The patients had no prior history of hypothermia before surgery. ③ The study focused on constructing a risk prediction model for hypothermia. ④ The study described model development, comparison, evaluation, and statistical methods. ⑤ The study type was cohort or case-control. ⑥ The language of publication was either Chinese or English.

Exclusion Criteria

① Full text or data was not accessible or incomplete. ② Duplicate publications. ③ Studies not using model development methods. ④ Publications in the form of titles, abstracts, letters, conference proceedings, or intervention drafts. ⑤ Non-Chinese or Non-English language publications.

Literature Screening and Data Extraction

EndNote 20 was used to remove duplicate records from the database search results. Two researchers with training in evidence-based nursing independently screened titles and abstracts based on the inclusion and exclusion criteria and read full texts to determine final inclusion. A standardized data extraction form, based on the Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies (CHARMS) checklist,²² was used to extract data from eligible studies. Extracted information included: author(s), publication year, study population, study design, measurement instruments, monitoring locations, frequency, candidate variables, methods for handling continuous variables, sample size, outcome event rates, missing data and handling methods, modeling methods, model performance, calibration methods, internal/external validation, number of model factors, and predictors.

Two researchers independently performed the literature screening and data extraction processes, with cross-validation. In case of disagreements, the issues were discussed or resolved through consultation by a third researcher via voting.

Quality Assessment of the Literature

The quality of the included studies was assessed using the Prediction Model Risk of Bias Assessment Tool (PROBAST).²³ This tool evaluates two main components: risk of bias and applicability. Two researchers independently evaluated the studies using the tool and cross-verified their findings. Discrepancies were discussed or resolved by a third researcher through voting.

Risk of Bias and Applicability Assessment

PROBAST provides a structured approach to identify potential reliability issues across four domains: participants, predictors, outcomes, and analysis, with 20 signaling questions. Each question is answered as “Yes” (Y), “Probably Yes” (PY), “No” (N), “Probably No” (PN), or “No Information” (NI). A “Yes” answer indicates lower reliability, while a “No” answer indicates higher reliability. “Probably Yes” (PY) and “Probably No” (PN) answers are allowed when there is insufficient information to be certain. A study is considered at low risk of bias if all signaling questions are answered as “Yes” or “Probably Yes.” If one or more questions are answered as “No” or “Probably No”, the study is considered at higher risk of bias. “No Information” (NI) means there is insufficient data, but this does not necessarily indicate bias.

The appli”, and “Unclear Concern.” If the research content (participants, predictors, outcomes) closely matches the systematic review question, it is deemed to have “Low Concern” regarding applicability. If there is a significant mismatch, it is rated as “High Concern.” If there is insufficient information to assess applicability, it is rated as “Unclear Concern.”

Results

Literature Screening Process and Results

A total of 240 relevant articles were initially identified. After removing duplicates using EndNote 20, 14 articles were selected based on the titles and abstracts, and after further screening, 8 articles were finally included.^24–31 The literature screening flow diagram is shown in Figure 2.

Figure 2 PRISMA study selection flow chart.

Basic Characteristics of the Included Studies

All 8 included studies developed risk prediction models for intraoperative hypothermia in total joint arthroplasty (TJA) patients. Among these, 3 studies used nomograms for modeling,^24,26,28 4 studies conducted external validation,^27–29,31 4 studies used risk formulae to calculate predictors,^25,27,30,31 and 1 study employed decision tree modeling.²⁹ Additionally, 4 studies were prospective cohort studies.^27–29,31 The basic characteristics of the included studies are detailed in Table 1.

Table 1 Characteristics of the Included Studies (n=8)

Basic Information on Model Development and Predictors

The total sample size for developing the 8 intraoperative hypothermia risk prediction models in THA patients ranged from 60 to 512 participants, with the event rate for the outcome ranging from 27.00% to 53.15%. Detailed information regarding the prediction models is provided in Table 2. Three studies reported missing data and their handling methods, with all missing data being excluded.^25,28,29

Table 2 Information of Hypothermia Risk Prediction Models for Total Jonit Arthroplasty (TJA) Patients

The predictors were classified into three categories: patient-related factors, surgical-related factors, and anesthesia-related factors, as detailed in Table 3. The number of predictors included in the models ranged from 3 to 6. The most common predictor was age (n=5). Among the patient-related factors, age and BMI were included in 62.5% and 37.5% of the studies, respectively. Surgical-related factors were included in the following proportions: intraoperative fluid volume (37.5%), surgical duration (25%), intraoperative blood loss (37.5%), operating room temperature (25%), intraoperative irrigation volume (25%), and whether active warming was used (25%). The anesthesia-related factors included: anesthesia duration (37.5%) and ASA classification (25%), as detailed in Table 4.

Table 3 Predictors of Models (n=8)

Table 4 Predictors for Repeat Reporting in Hypothermia Risk Prediction Models for TJA Patients

Model Validation and Performance

All 8 studies performed internal validation. Five of these studies used the Hosmer-Lemeshow goodness-of-fit test, with results showing P>0.05, suggesting no significant difference between predicted and observed values, indicating good model calibration. All 8 studies reported the area under the receiver operating characteristic curve (AUC),^24–31 with AUC ranging from 0.791 to 0.938, indicating good discriminatory ability of the models. Internal validation was conducted using the following methods: two studies used Bootstrap resampling,^24,26 three studies used random splitting for validation,^27–29 and three studies did not specify the method of internal validation.^25,30,31 Five studies performed external validation.^{25,27–29,31} Seven studies reported sensitivity and specificity,^{24,25,27–31} suggesting that the models had good diagnostic ability, as shown in Table 2.

Quality Assessment Results of the Included Studies

The quality of the prediction model studies was assessed using PROBAST. The assessment results indicated that 7 studies had a high overall risk of bias.^{24–26,28−31} In the participant domain, 4 studies were prospective,^27–29,31 while the other 4 studies were retrospective,^24–26,30 which might have resulted in missing data and known disease outcomes, leading to a high risk of bias in this domain. In the statistical analysis domain, 3 studies deleted data without specifying the reasons for or methods of data handling,^25,28,29 which may have introduced bias in the results. Three studies used only random splitting for internal validation,^27–29 and three studies did not specify the internal validation method,^25,30,31 leading to concerns about the comprehensiveness of model performance and fitting. Regarding applicability, the study by Zhou Yu et al²⁴ included patients with other types of joint arthroplasty, which led to a high concern for applicability, as detailed in Table 5.

Table 5 Bias Risk and Applicability Evaluation of Hypothermia Risk Prediction Model for TJA Patients (n=8)

Discussion

Analysis of the High Risk of Bias and Applicability in Low-Temperature Risk Prediction Tools for Joint Arthroplasty Patients

The models included in this study showed a high risk of bias in both the participant and data analysis domains. The reasons for this bias are as follows:

1. Participant Domain: Among the 8 studies included in this review, 4 were prospective studies,^27–29,31 and 4 were retrospective studies.^24–26,30 In these cases, the inclusion of patients with missing data and known disease outcomes, along with the use of historical clinical data to construct prediction models, may lead to incomplete or suboptimal data quality,³² which in turn increases the risk of bias. The predominance of retrospective designs among the included studies poses a significant challenge to clinical reliability. Such methodological limitations may lead to an overestimation of model performance due to inherent selection bias and the use of historical data. In clinical practice, this implies that models might underperform when faced with real-time, prospective patient data, potentially leading to inaccurate risk stratification by nursing staff.
2. Statistical Analysis Domain: Three studies directly deleted missing data without providing explanations or clarifications of the procedures,^27–29 which may cause bias in the results.³³ The practice of excluding participants with missing data, observed in several studies, directly impacts clinical generalizability. This approach often inadvertently filters out the most complex or critically ill patients who frequently have incomplete records. Consequently, the resulting models may lack the robustness required to guide decision-making for high-risk surgical candidates, where accurate hypothermia prediction is most vital. Furthermore, 3 studies only used random data splitting for internal validation,^27–29 and 3 studies did not specify their internal validation methods.^25,30,31 This indicates potential issues with model performance and fitting. Additionally, 4 studies did not conduct external validation of the prediction models,^24,26,29,30 which may result in biased performance assessments and an increased risk of overfitting. Single-center study data inherently carry the risk of selection bias, and only 50% of the studies conducted external validation, which limits the generalizability of the models.
3. Applicability Evaluation: Seven studies overall showed low-risk applicability, indicating that the models included in the studies are theoretically applicable. However, 4 studies used tympanic temperature measurement,^24,27,28,30 while 3 studies used nasopharyngeal temperature measurement,^25,26,29 and 1 study used infrared forehead thermometry.³¹ The inconsistency in measurement methods may have affected data collection results, leading to potential biases. Moreover, infrared forehead temperature measurement is not the recommended method in guidelines and consensus, which could cause discrepancies in monitoring hypothermia and potentially affect study outcomes.

Future research should focus on enhancing the scientific rigor and standardization of these tools. In the process of developing risk prediction models, not only should sample size be considered, but multi-center data applications should also be prioritized to improve model generalizability. Missing data should be handled using single or multiple imputation methods to reduce the impact of data loss on statistical analyses.³⁴ Additionally, expert opinions or consensus should be incorporated when selecting predictors, providing direction for future predictor selection. Internal and external validation of the models should be conducted, utilizing independent datasets for external validation, in order to evaluate model performance from multiple dimensions and minimize the risks of overfitting or underfitting. Regarding applicability, the study by Zhou Yu et al included patients with other types of joint arthroplasty, which led to a high concern for applicability, as detailed in Table 5. The predominance of retrospective designs among these studies poses a significant challenge to clinical reliability. Such methodological limitations may lead to an overestimation of model performance due to inherent selection bias. In practical nursing, this implies that models might underperform when faced with real-time, prospective patient data, potentially leading to inaccurate risk stratification and delayed interventions.

The Need for Improvement in Low-Temperature Risk Prediction Models for Joint Arthroplasty Patients

Single and Limited Modeling Methods

In the 8 risk prediction models included in this study, 7 used traditional Logistic regression to construct the prediction models, while 1 study employed machine learning methods. Among these, 7 models had an AUC > 0.8, suggesting good overall performance. However, the bias risk assessment results show that the included studies had a high risk of bias, particularly in the participant and data analysis domains. In terms of applicability, the evaluation showed low-risk applicability for most models. Although some studies used Random Forest modeling, which can capture non-linear relationships and is not affected by multicollinearity, the Random Forest model showed an AUC greater than 0.8, indicating good discriminatory ability, with sensitivity (0.7) and specificity (0.79) suggesting the model can effectively identify most positive outcomes and distinguish negative cases. However, because external validation was not performed, the model’s generalizability remains unconfirmed. Future studies using machine learning methods should ensure that after fitting the internal data, the model’s generalizability is validated across different centers and populations. Expanding the sample size and collecting multi-center data will improve internal and external validation, thus enhancing model clinical applicability and stability.

Differences in Research Design and Data Quality

Among the included studies, 4 were retrospective cohort studies,^24–26,30 and 4 were prospective cohort studies.^27–29,31 Half of the studies used historical clinical data for model development, which could lead to potential issues with data quality.³² Additionally, all studies were single-center, limiting their generalizability due to population and environment constraints. Three studies reported handling of missing data,^25,28,29 but the approach was limited to direct deletion, which may introduce bias in model performance and increase the risk of overfitting. Other studies did not specify how missing data were handled, which may have contributed to statistical biases. Future research should adopt imputation or filling methods to standardize the handling of missing data.³⁵

Variability in Validation Methods and Completeness

Zhao et al conducted external validation with a group of 206 patients,²⁸ where the incidence of intraoperative hypothermia (IH) was 48.06%, close to the 53.16% incidence in the modeling group. The sensitivity, specificity, and accuracy of the external validation group were all close to 90%, indicating good stability and accuracy of the model in an independent sample. Li et al and Li also conducted external validation with accuracies exceeding 80%.^27,31 Liu Xiaohui et al reported good external validation,²⁵ but the accuracy was only 76.7%, and the sample size for external validation was small, which may reduce the stability of the validation indicators. Increasing the sample size for further validation would help improve the generalizability of the model. Analysis of the proportion of external validations indicates that there is still room for improvement in the validation processes. Comprehensive validation of prediction models can enhance their strengths and performance, reduce impacts on applicability and accuracy, and improve research quality.

Differences in Core Temperature Monitoring Methods

Intraoperative temperature monitoring should be continuous and dynamic, which is essential for perioperative temperature management and can help reduce postoperative complications. Core temperature monitoring methods vary in accuracy. Blood temperature measurement is the most accurate but is costly and not widely applicable. In contrast, esophageal or nasopharyngeal temperature measurement is currently the most recommended and practical method.¹ In Li et al’s study,³¹ the researchers used infrared forehead temperature monitoring, which is not a recommended core temperature monitoring site. This could lead to bias in the temperature monitoring results and may affect model performance. It is also important to ensure that the probe is placed at the correct depth and site to minimize abnormal temperature measurements.

Differences in Predictor Selection

Upon integrating the final predictors across the included studies, it was found that age (n=5), BMI (n=3), intraoperative fluid volume (n=3), anesthesia time (n=3), intraoperative blood loss (n=3), and hemoglobin (n=3) were the most frequently used predictors in low-temperature risk prediction models for joint arthroplasty patients. However, there were differences in the number and types of predictors included in the models, with some models not including highly correlated predictors. This lack of coverage of key predictors could affect the accuracy of model predictions.

Currently, low-temperature risk prediction models for joint arthroplasty patients are still in their early stages. Future research should focus on standardizing model development and validation, constructing large-scale multi-center models, and designing studies in line with the PROBAST criteria to provide more applicable prediction tools for clinical practice.³⁵

Limitations of Single-Center Studies in Predicting Hypothermia Risk in Total Joint Arthroplasty Patients

Current research on hypothermia risk prediction models for total joint arthroplasty (TJA) patients has predominantly been conducted in China, with most adopting single-center designs. This characteristic leads to significant limitations in model generalizability, representativeness, and clinical applicability. These limitations can be analyzed from four perspectives: differences in healthcare systems, population-specific characteristics, research design constraints, and heterogeneity in clinical practice.

1. Differences in Healthcare Systems: Chinese medical institutions have gradually developed standardized preoperative assessments, intraoperative procedures, and temperature monitoring protocols.¹ However, compared to developed countries like those in Europe and North America, hypothermia management in China started relatively late, resulting in differences in measurement methods and timing. For example, the AORN recommends recording core body temperature every 30 minutes during surgery,³⁶ while the models included in this study monitored core temperature every 10–15 minutes. The frequency of monitoring may only be suitable for Chinese settings and applying these models directly to European or American contexts may introduce biases due to differences in the granularity of monitoring data. Additionally, the disparity in intraoperative temperature management between tertiary hospitals and primary-level hospitals in China, with only 48.84% of tertiary hospitals having established emergency protocols for hypothermia,³⁷ further narrows the model’s applicability in domestic settings. Furthermore, key predictive factors such as “surgical duration” and “intraoperative fluid volume” (each contributing 37.5% to the model as seen in Table 4) are based on data from single-center studies in China. For example, in Bin Zhao et al’s model,²⁸ a threshold of intraoperative fluid volume greater than 1500mL is defined as high risk, whereas in European and American studies, the corresponding threshold is typically set at 2000mL.⁵ Directly applying the Chinese model could lead to incorrect risk assessments for patients in these regions, further reducing the model’s cross-system adaptability.
2. Differences in Population Characteristics: The models included in this study were all based on Chinese populations, which exhibit significant regional differences, limiting the representativeness of the models. The average BMI in China (24.7 kg/m²) is lower than that in Europe and North America (28.3 kg/m²).⁵ Literature reviews and predictive modeling analyses indicate that BMI is a critical predictor of hypothermia risk.^36,38,39 In China, a BMI of less than 18.5 kg/m² is considered high risk, but this standard may be less applicable to populations in Western countries, where a higher baseline BMI is common. For instance, in Leilei Li et al’s model,³¹ the incidence of hypothermia for patients with a BMI below 18.5 kg/m² was 58.2%, while studies from Western countries report a 50% incidence for those with a BMI below 22 kg/m.²⁵ These physiological differences directly affect the applicability of predictive thresholds. The latest data from the National Hip Fracture Database (NHFD) in the UK shows an increasing incidence of hip fractures, with 72,160 cases reported in 2023.⁴⁰ Meanwhile, China’s annual volume of joint arthroplasties reaches 577,000,⁴¹ suggesting a higher volume of patients and more granular data for research. However, there are also differences in the underlying diseases and medication usage between Chinese and Western populations,⁶ which can impact factors like vascular tone and metabolism, indirectly influencing hypothermia risk. Current models have not accounted for the control of comorbid conditions or the use of anticoagulants, which may reduce the accuracy of predictions in different populations. For example, Zhou Yu et al’s study found that a hemoglobin level below 100 g/L indicated high risk in Chinese patients,²⁴ whereas in Western populations, anticoagulated patients with a hemoglobin level below 90 g/L face significantly higher risks.⁶ These medication and comorbidity differences further limit the model’s generalizability. The ethnic homogeneity of the current evidence base necessitates cautious clinical implementation. Given the physiological variations in BMI and hemoglobin thresholds between Chinese and Western populations, these models require local recalibration before cross-regional adoption. Failure to account for these population-specific variables could result in misaligned nursing interventions and compromised patient safety in diverse clinical settings.
3. Research Design Limitations: Single-center study designs inherently carry risks of selection bias, environmental interference, and data homogeneity, which undermine the stability and external validity of the models. For example, some studies included only ASA I–II patients,²⁸ excluding those with higher ASA classifications, potentially diminishing the model’s ability to predict risks in critically ill patients. China’s vast geographic diversity also introduces climate differences that may impact temperature management practices. Additionally, single-center studies often rely on unique temperature monitoring methods, with variations in the measurement site and timing. This lack of standardization in core temperature measurement as a predictive factor reduces the consistency of the models. Moreover, differences in blood loss estimation methods among studies can exacerbate data bias, further weakening the model’s reliability.
4. Heterogeneity in Clinical Practices: The models currently developed are based on clinical practices in China, where surgical techniques, equipment configurations, and nursing procedures significantly differ from those in other regions. For instance, the use of minimally invasive techniques in joint replacement surgery can reduce surgical time and tissue trauma. In the United States, the proportion of day-case hip and knee replacements increased from 2010 to 2017, but this practice is still not widespread in China.⁴² The efficient completion of same-day surgeries in the U.S. relies on coordinated surgical operations and integrated post-operative monitoring, which are still in the early stages of implementation in China. These differences further limit the clinical applicability of the models.

Future Directions for Hypothermia Risk Prediction Models in Total Joint Arthroplasty Patients

Prevention and management of hypothermia are key to improving perioperative nursing safety and quality. However, current domestic protocols for hypothermia prevention and clinical assessment are still lacking, with only 48.84% of tertiary hospitals having established emergency protocols for hypothermia.⁷ In addition to temperature monitoring and various warming measures, it is crucial to screen high-risk individuals for influencing factors, establish preoperative hypothermia assessments, and develop appropriate risk prediction models to guide clinical evaluation.

Model Optimization Directions

Promoting Multicenter Data Sharing and Collaboration: All eight studies included in this research were single-center studies, which may introduce selection bias and limit the generalizability of the models. Future research should be led by top nursing teams and collaborate with various hospitals across regions and hospital levels to create a shared database for hypothermia risk prediction in TJA patients. One study has proposed the use of mobile apps to calculate perioperative hypothermia risk scores through multicenter, prospective, observational cohort studies across more than 30 hospitals. This approach could enhance data standardization and allow for better model optimization. By incorporating multicenter data to improve model parameters and performing external validation using independent cohorts, model optimization could be better achieved. However, careful attention must be paid to whether the model’s application scenario aligns with the population and clinical settings in other regions.⁴³

Artificial Intelligence Leadership: With the development of artificial intelligence (AI) technologies in recent years, machine learning (ML) has increasingly been applied to clinical research, including in disease progression prediction,^13–16 decision support,^17–19 and risk prediction.^19–21 Intraoperative hypothermia is a common complication in TJA, closely linked to factors such as long surgical duration, anesthesia methods, and age-related decline in temperature regulation. While current preventive measures have shown some effectiveness, their implementation rate is low, and they lack precise risk stratification and personalized interventions. Recent advances in machine learning have shown great potential for handling multidimensional and nonlinear relationships, which is an advantage in predicting hypothermia risk. The strengths of ML lie in its ability to iteratively test complex relationships between numerous potential factors and utilize ensemble algorithms for high predictive accuracy and precision. Compared to traditional statistical methods, ML automatically identifies potential relationships between variables and builds more accurate, efficient models that align with clinical needs.⁴⁴ Therefore, clinical nursing research should closely integrate big data and AI technologies, exploring other machine learning methods for model construction in future studies.

Conclusion

This study indicates that while IH risk prediction models for TJA patients show promise, they currently exhibit a high risk of bias and are still in their infancy. To bridge the gap between model development and clinical utility, future research must prioritize rigorous external validation using large-scale, multicenter datasets to mitigate the risk of overfitting. Furthermore, transitioning these validated algorithms into visualized, point-of-care clinical decision support tools will be instrumental in translating predictive insights into improved thermoregulation outcomes and enhanced patient safety for TJA patients. Future nursing efforts should focus on the triad of “model optimization – clinical application – scientific education”. Through multicenter collaboration and standardized application, the quality of models can be improved, internal and external validation can be strengthened to enhance generalizability, and the models can be transformed into clinical practice to reduce the incidence of hypothermia, improve patient outcomes, and enhance patient satisfaction.