Local hyperthermia therapy for the pain management of knee osteoarthritis: a network meta-analysis of randomized controlled trials

Abstract

Purpose

• The aim of this study was to evaluate the efficacy of local hyperthermia therapies (LHTs) and their combination with exercise for pain improvements in adults with knee osteoarthritis (OA).

Methods

• A systematic search was conducted across PubMed, Web of Science, Scopus, EMBASE, Cochrane Central Register of Controlled Trials, and PEDro (up to May 25, 2026) to identify randomized controlled trials (RCTs) investigating thermotherapy versus control or no interventions in adults with knee OA. A random-effects network meta-analysis was performed. Pain outcomes were assessed for immediate change from baseline (≤1 week) and short-term change (≥4 weeks and <12 weeks/3 months). Methodological quality was evaluated using the PEDro scale. Certainty of the evidence was graded using the Confidence in Network Meta-Analysis (CINeMA) framework. Standardized mean differences (SMDs) with 95% confidence intervals (CIs) were calculated.

Results

• A total of 73 RCTs (5,270 participants) were included, evaluating interventions such as high-intensity laser therapy (HILT), ultrasound, diathermy, and hot pack. Associations with the greatest decrease were observed for HILT with exercise (SMD: −3.49, 95% CI: −4.65 to −2.32) in immediate pain and for organic hot pack with exercise (SMD: −3.93, 95% CI: −5.65 to −2.21) in short-term pain. These interventions combined with exercise ranked highest for efficacy as reflected by the surface under the cumulative ranking curve (SUCRA) values.

Conclusions

• LHTs, combined with exercise, are likely effective modalities for improving immediate and short-term pain in knee OA. Due to low certainty of evidence, larger high-quality RCTs are required to confirm these synergistic effects.

Introduction

Osteoarthritis (OA) is the most prevalent form of arthritis, predominantly affecting the knee joint (1). Current therapeutic approaches for knee OA primarily focus on pain management during daily activities. While nonsteroidal anti-inflammatory drugs (NSAIDs) offer significant immediate benefits for knee OA, their high costs and potential adverse effects restrict their use in certain real-world contexts (2, 3, 4).

Non-pharmacological interventions are recommended in clinical guidelines, among which local hyperthermia therapy (LHT) is one of the most widely recommended modalities for reducing pain and stiffness, as well as for improving mobility (5, 6). Techniques for hyperthermia therapy include superficial hot compress, photothermal therapy, and electromagnetic diathermy, and these techniques may also be combined with exercise or other physical therapy interventions (7). Currently, a substantial number of randomized controlled trials (RCTs) have investigated the use of LHTs for the management of OA. However, a high-quality RCT appears to demonstrate only a modest beneficial effect in the management of hand OA, and this effect may have been overestimated (8). Moreover, although relevant meta-analyses have confirmed the efficacy of various LHT applications (9, 10, 11), it remains unclear whether differences in efficacy exist across different heating modalities and to what extent therapeutic benefits may be influenced by factors such as heat penetration efficiency and penetration depth (12). Furthermore, there is uncertainty whether a combination of exercise and local hyperthermia interventions yields greater improvements in pain intensity among adults with knee OA, when compared with either treatment alone.

To address these gaps, we retrieved and reviewed all pertinent RCTs related to heat application and conducted a network meta-analysis (NMA) of the available evidence. This study seeks to establish a more comprehensive foundation for clinical practice by comparing all thermal agents in the treatment of adult patients with knee OA.

Materials and methods

Study design and protocol registration

The present study was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) (13). The protocol of this systematic review was registered in PROSPERO (registration number: CRD42024587728).

Literature search strategy

An online systematic search was conducted for eligible trials across electronic databases, including MEDLINE (PubMed), Web of Science, Scopus, EMBASE, Cochrane Central Register of Controlled Trials, and PEDro (https://www.pedro.org.au) from their inception until December 25, 2024. Furthermore, an updated search of these databases was performed up to May 25, 2026. The search strategy integrated terminology related to knee OA, RCT, and various thermal intervention agents, with detailed methodology available in Supplementary Table 1A, B, C, D, E, F (see section on Supplementary materials given at the end of the article). Additionally, manual searches and citation tracking, both backward and forward, were performed utilizing reference lists from other systematic reviews and RCT articles.

Eligibility criteria

Eligible studies encompassed RCTs of non-pharmacological treatments that assessed a single type of LHT, either as a stand-alone treatment or adjunctively, compared with an active or true control in adult patients with knee OA. LHT was defined as a therapeutic modality that induces temporary hyperthermia within the knee tissues and/or the superficial tissues of the knee through thermal physical agents. Unlike recent reviews, we excluded studies that employed multiple hyperthermic agents or interventions with incidental thermal effects where hyperthermia was not the primary therapeutic mechanism (e.g. low-intensity laser therapy). Studies utilizing an active control design were deemed eligible provided that the treatment protocol involved noninvasive and athermal applications. Trials were considered eligible if at least 80% of the enrolled patients had knee OA. To reduce the impact of small-study effects and enhance the robustness of pooled estimates, studies with more than 10 participants per group were included. Moreover, only full papers published in peer-reviewed journals were accepted. Characteristics of participants for eligibility criteria are listed in the study protocol material (https://www.crd.york.ac.uk/PROSPERO2/view/CRD42024587728).

Study selection

In the process of selecting studies, duplicate records were eliminated, followed by independent screening of titles and abstracts. Two researchers (R Z and H G) independently conducted the study selection process, with discrepancies resolved through discussion or, if necessary, the involvement of a third reviewer (Z L). Colleagues were consulted to assist with the translation and evaluation of articles published in languages other than English.

Data extraction

The extracted data encompassed details of clinical trial reports, including authors, their region or country, and year of publication. Patient information collected comprised age, sample size, gender, body mass index (BMI), Kellgren and Lawrence (K/L) grade, intervention administration, outcome assessment time point, and outcome questionnaire of interest. Particular types of interventions were classified according to group names designated by the authors and the definitions provided in Supplementary Table 2.

The primary and secondary outcomes were the mean changes in knee pain from baseline to the end point and those during short-term follow-ups. The preferred outcome measure was the pain subscale of the Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC) (14). In accordance with the literature recommendations, when outcomes in the included RCTs were measured using different scales, the prioritization of these outcomes adhered to the methodology outlined in Supplementary Table 3 (15). For each outcome, the changes from baseline at the immediate time point (≤1 week) and short-term time point (≥4 weeks and <12 weeks/3 months) were extracted. In instances where multiple follow-up points were reported within a single timeframe, the longest follow-up was utilized for analysis. No additional data were requested from the authors.

Quality review

The PEDro scale was utilized to assess the methodological quality of each RCT (16), independently by two researchers (R Z and H G). Any discrepancies were resolved through consultation with a third researcher (Z L). The PEDro scale consists of 11 items. The final PEDro score for each trial was derived by summing the ratings for items 2–11. According to the resultant PEDro score, the quality of methodology was classified as excellent (range: 9–10), good (range: 6–8), fair (range: 4–5), or poor (range: 0–3).

Data synthesis and analysis

The synthesis and analysis of data were conducted by L Y. The means and standard deviations (SDs) or standard errors were directly obtained from the presented tables, while graphical data were extracted using WebPlotDigitizer (version 4.5.0). According to the Cochrane Handbook (https://www.training.cochrane.org/handbook), when studies were reverse scaled (i.e. higher values indicated better outcomes rather than lower values), the mean in each group was multiplied by −1. Based on the Cochrane Handbook, for studies that reported mean and standard error of the mean (SEM), the SD was calculated using the following formula: SD = SEM*sqrt(N). For studies reporting median and interquartile range, the mean and SD were calculated using an online tool (https://www.math.hkbu.edu.hk/∼tongt/papers/median2mean.html). Change values were extracted directly whenever the means and SDs of the pre- and post-intervention scores were available (17). If data concerning the percentage difference from baseline (i.e. prior to treatment) were provided, they were incorporated directly into the meta-analysis. When studies did not report mean changes or SDs, these values were imputed or calculated based on standard procedures (18, 19). Data from multiple groups were aggregated using a well-established method for trials with several similar interventions and comparable effect sizes (20). For trials employing a crossover design, data could be subdivided into multiple comparable groups.

We performed a frequentist NMA within a graph-theoretical framework to simultaneously synthesize direct and indirect evidence across all eligible comparisons, including sham-controlled, placebo-controlled, and active-controlled trials. A random-effects model was specified a priori to account for anticipated between-study heterogeneity arising from variations in population characteristics, intervention protocols, and follow-up durations. Hedges’ g, a small-sample bias-corrected variant of the standardized mean difference (SMD), was selected as the pooled effect metric, with all results presented as point estimates accompanied by 95% confidence intervals (CIs). To interpret the magnitude of the effect, SMD values were classified as follows (21): 0.2 indicates a small effect size, 0.5 a moderate effect size, and 0.8 a large effect size. For pairwise comparisons generated from the random-effects NMA, statistical superiority was based on the 95% CI of the Hedges’ g estimate. The intervention hierarchy was also ranked using the surface under the cumulative ranking curve (SUCRA) values (22). Heterogeneity was assessed utilizing the I² statistic. Local inconsistency between direct and indirect comparisons was evaluated using the node-splitting method (23). Finally, publication bias and risks of small-study effects were examined through the comparison-adjusted Egger’s linear regression test to assess funnel plot asymmetry (24). To ensure the robustness of primary findings and the interpretability of effect estimates, we performed separate NMAs as sensitivity analyses, for the studies using the WOMAC pain subscale. The subscale mean differences (MDs) were normalized to a scale from 0 to 100, allowing direct interpretation against published minimal clinically important differences (MCIDs). All analyses were conducted using R software (version 4.5.3) with the ‘netmeta’ package (version 3.6-0) (25) and Free Statistics analysis platform (version 2.1.1). The estimated effect was considered statistically significant if its 95% CI did not include 0.

Credibility assessment

The Confidence in Network Meta-Analysis (CINeMA) tool was utilized to evaluate the confidence rating of each comparison (26). The CINeMA framework consists of six domains, such as within-study bias, reporting bias, indirectness, imprecision, heterogeneity, and incoherence. Based on the contributions from those domains, the certainty of evidence was graded as high, moderate, low, or very low.

Results

Study characteristics

A total of 6,966 citations were identified through database searches and other sources, and 2,827 duplicate records were eliminated. Of the remaining 4,139 studies, 3,914 were excluded based on their titles and/or abstracts. Full-text screening for eligibility was conducted on 225 articles, resulting in the inclusion of 73 articles, as illustrated in the PRISMA flow diagram (Fig. 1), along with documented reasons for exclusion.

Figure 1.

PRISMA flow diagram of the search process. PRISMA, Preferred Reporting Items for Systematic Reviews and Meta-Analyses; RCT, randomized controlled trial.

The demographics and study characteristics of the included RCTs are summarized in Supplementary Table 4 (27, 28, 29, 30,31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99). The corresponding dosages and parameters in each trial are detailed in Supplementary Table 5. There were 24 placebo-controlled studies, with 8 of these incorporating exercise as an adjunct intervention. Of the enrolled RCTs, 57 employed a two-arm design, while the remaining 16 were multi-arm trials. Only one study utilized a crossover design and was consequently divided into two RCTs for analysis. The publication years of these studies ranged from 2004 to 2026. Regarding the study regions, Turkey and China were the most frequently represented, collectively constituting 43.8% of all study regions. The total number of treatment sessions varied from 5 to 42. Concerning the time points of outcome assessments, all 73 studies reported pre- and post-treatment pain outcomes, and 40 studies also conducted short-term pain follow-ups.

In the presented NMA, all these trials involved 5,270 patients (primarily between 50 and 70 years of age), with 75.7% being female according to reports from 57 trials. The patients’ BMI ranged from normal to obese. A radiographic diagnosis of K/L grades 2–3 was observed in 42.1% of the patients, based on data from 54 studies. For the design of interventions, eight types of thermal interventions were considered, including high-intensity laser therapy (HILT), continuous ultrasound (US), electromagnetic diathermy, infrared radiation, moxibustion, hot pack, organic hot pack, and steam hot pack.

Quality assessment

Among 73 studies, PEDro scores for 39 were extracted from the PEDro database and those for the remaining 34 were independently assessed by two researchers (R Z and H G). The mean (SD) PEDro score was 6.93 (1.48). Of these, 14 studies were rated as excellent, 44 as good, 15 as fair, and none as poor. A total of 71 studies appropriately implemented random allocation. One item (concealed allocation) received positive scores in 42 studies, whereas another item (blind therapists) was positively scored in only 12 studies. Additionally, 41 studies were scored positively on another item (blind assessor). Overall, there were identified risks of bias concerning selection, performance, and detection across the studies. The detailed methodology quality of the included papers, as assessed by the PEDro classification scale, is presented in Supplementary Table 6.

Primary outcomes

A total of 21 interventions were conducted for immediate pain assessment. Among the 73 studies reviewed, the most common outcome measure for pain assessment was the WOMAC in 44 studies, followed by the visual analog scale (VAS) in 22 studies, the 36-Item Short Form Survey (SF-36) in 4 studies, the Knee injury and Osteoarthritis Outcome Score (KOOS) in 2 studies, and the numeric rating scale (NRS) in 1 study. The network plot for immediate pain is shown in Supplementary Fig. 1. When stand-alone LHTs were analyzed, the most pronounced pain decrease was associated with HILT (SMD: −2.47, 95% CI: −4.32 to −0.62), followed by organic hot pack (SMD: −1.53, 95% CI: −2.29 to −0.76) and infrared radiation (SMD: −1.38, 95% CI: −2.39 to −0.38) (Fig. 2). In terms of combined LHTs, HILT plus exercise (SMD: −3.49, 95% CI: −4.65 to −2.32), continuous US plus exercise (SMD: −3.34, 95% CI: −4.57 to −2.11), and organic hot pack plus exercise (SMD: −3.13, 95% CI: −4.24 to −2.03) demonstrated greater efficacy in pain reduction compared with other approaches. Considering all thermal interventions, HILT plus exercise exhibited the highest probability of being the most effective treatment for immediate pain relief based on the SUCRA value of 0.96. As expected, the efficacy of single interventions was generally inferior to that of their combinations with exercise. However, it should be noted that combined interventions with better efficacy were supported by low-level evidence (Supplementary Table 7).

Figure 2.

Forest plot for the immediate effects of treatments on knee pain. CI, confidence interval; HILT, high-intensity laser therapy; PT, physical therapy; SMD, standardized mean difference; SUCRA, surface under the cumulative ranking curve; US, ultrasound.

In pairwise comparisons within the network of all trials (Supplementary Table 8), HILT plus exercise was associated with improved pain compared with most individual LHTs (e.g. different kinds of hot packs). Continuous US was associated with less improvement in pain compared with most individual LHTs, whereas continuous US plus exercise provided greater improvement than the remaining combined interventions.

Secondary outcomes

The network plot for short-term pain, derived from 40 RCTs covering 15 thermal interventions, is shown in Supplementary Fig. 2. In comparison with placebo, combined LHTs, such as organic hot pack plus exercise (SMD: −3.93, 95% CI: −5.65 to −2.21), hot pack plus exercise (SMD: −3.67, 95% CI: −5.07 to −2.27), and continuous US plus exercise (SMD: −2.93, 95% CI: −4.63 to −1.23), demonstrated greater efficacy in pain reduction (Fig. 3). Among individual LHTs, organic hot pack (SMD: −1.61, 95% CI: −2.54 to −0.69) was significantly associated with reductions in pain over short-term periods, followed by diathermy (SMD: −1.00, 95% CI: −1.71 to −0.30) and moxibustion (SMD: −0.83, 95% CI: −1.34 to −0.32). When considering all thermal trials, the combination of organic hot pack and exercise had the highest likelihood of being the most effective treatment for short-term pain relief, with a SUCRA value of 0.93. Similarly, LHTs combined with exercise also outperformed single interventions in terms of short-term pain relief, despite downgrading confidence (Supplementary Table 9).

Figure 3.

Forest plot for the short-term effects of treatments on knee pain. CI, confidence interval; HILT, high-intensity laser therapy; PT, physical therapy; SMD, standardized mean difference; SUCRA, surface under the cumulative ranking curve; US, ultrasound.

Organic hot pack plus exercise was also associated with improved short-term pain in pairwise comparisons against several treatments included in the analysis (Supplementary Table 10), notably most individual LHTs (e.g. diathermy and moxibustion). Hot pack plus exercise achieved the next highest SUCRA value, with a larger sample size than organic hot pack plus exercise.

Sensitivity analyses

In relation to the reduction in immediate pain, the network geometry for the trials using the WOMAC pain subscale remained intact (Supplementary Fig. 3). The NMA using the WOMAC MD on a normalized scale from 0 to 100 revealed that HILT plus exercise was consistently associated with a significant decrease in pain (MD: −20.83, 95% CI: −36.88 to −4.77), whereas continuous US plus exercise resulted in the most pronounced reduction (MD: −28.76, 95% CI: −45.74 to −11.78) (Supplementary Table 11). In addition, there was a significant association of organic hot pack with greater pain improvement, among individual LHTs (MD: −16.67, 95% CI: −27.51 to −5.83).

For the reduction in short-term pain, the network structure based on the MDs from the WOMAC pain subscale was mostly preserved, despite the exclusion of two nodes, including organic hot pack plus exercise and hot pack plus exercise (Supplementary Fig. 4). Diathermy plus exercise demonstrated the most substantial enhancement in pain relief (MD: −33.69, 95% CI: −47.98 to −19.40) (Supplementary Table 12). Among individual LHTs, organic hot pack was consistent with the main results (MD: −17.42, 95% CI: −27.02 to −7.81).

The node-splitting results revealed minimal inconsistencies between the direct and indirect evidence; these findings were corroborated by visual inspection of forest plots (Supplementary Figs 5 and 6). Comparison funnel plots for pain improvement exhibited a roughly symmetrical distribution (Supplementary Figs 7 and 8). Combined with the results of Egger’s linear regression test (immediate pain: P = 0.837; short-term pain: P = 0.693), there was no evidence of significant publication bias or small-study effects in the NMA.

Discussion

To our knowledge, this NMA, conducted with strict inclusion criteria, constitutes the first review to assess the efficacy of various types of LHT for knee OA. The analysis was conducted based on 73 RCTs including over 5,000 participants randomly assigned to 21 different arms: non-pharmacological (physiotherapy) interventions, sham controls, or placebo controls. The results demonstrated that most forms of LHTs were superior to placebo for immediate pain relief (SMD < −0.2), whether applied as an adjunct to exercise or independently. In short‑term follow‑ups, LHTs maintained a directional trend mirroring that observed in the immediate effects. These findings are consistent with the 2019 American College of Rheumatology recommendations, which conditionally endorse thermal modalities as a non-pharmacological intervention for pain management in knee OA (100).

Clinical implications

The application of thermal effects in disease treatment has gained widespread acceptance (101). In managing OA manifestations, it is well known that hyperthermia enhances tissue metabolism, improves collagen elasticity, augments capillary blood flow, and alleviates muscle spasms (102). Additionally, studies have demonstrated that thermal exposure reduces distal motor and sensory latencies, thereby increasing nerve conduction velocity, elevating pain thresholds, and exerting analgesic effects (103). According to modes of heat delivery, LHT can be classified into two distinct types: superficial and deep heating. Superficial heating modalities, such as hot packs, primarily warm the skin and subcutaneous tissues (104). In contrast, deep-heating agents can treat larger areas and greater tissue volumes, potentially yielding superior therapeutic outcomes for deep tissue treatment (105, 106).

NMA is acknowledged as a source of high-grade clinical evidence. Its strength lies in its ability to enable both direct and indirect comparisons within a network of RCTs. This comprehensive methodology permits a more robust and unbiased assessment of intervention effectiveness in a clinical context. The majority of studies have demonstrated that various physiotherapeutic interventions possess comparable efficacy. Our findings offered a thorough comparison of the efficacy of different physiotherapeutic interventions and thermal modalities. Regarding superficial heating applications, hot packs were the most extensively used LHTs for knee OA. We observed that they were associated with a greater effect on pain management compared with placebo. Similarly, the application of organic hot pack was consistently associated with significant clinical benefits for pain relief, both immediately after use and across short-term follow-up periods. These results remained basically robust when evaluated using the WOMAC pain subscale. In our findings, multiple interventions, such as continuous US plus exercise, far exceeded an MCID of approximately 9 points for the WOMAC pain subscale (107, 108), supporting their clinical relevance. For deep hyperthermia therapies used in the management of musculoskeletal disorders, we found that these were associated with significant improvements in immediate pain. For instance, HILT showed statistically significant superiority over all other stand-alone modalities. Previous investigations have demonstrated that, in patients with knee OA, SUCRA values favor deep heat therapies, owing to their advantageous thermal penetration capabilities over superficial heat applications (109). Consistent with this observation, we hypothesize that heat penetration efficiency and depth are key factors influencing the efficacy of hyperthermia application.

As reported in previous reviews, exercise contributes to pain alleviation (110). Similarly, we found that patients with knee OA benefited from exercise, for both immediate and short-term pain relief. Moreover, exercise generally outperformed single thermal modalities (e.g. hot pack and diathermy) across both time points. However, our NMA consistently indicated that integrating thermal modalities with physical activity yielded the most substantial and sustained clinical benefits, ranking highest in our analysis. This potential additive effect may be attributed to thermal modulation increasing local blood flow, thereby enhancing nutrient supply and waste removal in OA joints. This synergistic interaction with exercise further improves muscle oxygenation and mitochondrial function, reducing lactic acid accumulation that otherwise exacerbates pain. Based on SUCRA values, HILT plus exercise emerged as the optimal intervention for immediate pain control, followed by continuous US plus exercise and organic hot pack plus exercise. Over short-term follow-ups, the combination of organic hot pack and exercise demonstrated the highest overall efficacy in pain reduction among all treatment arms, followed by hot pack plus exercise. These discrepancies in therapeutic efficacy may be attributed to the inherent properties of the thermal agents used.

Limitations

Our analysis had some limitations. First, the thermal effect of physiotherapy is contingent upon the elevation of tissue temperature. Some studies explicitly reported the induction of thermal sensations, whereas others only reported disparate equipment settings. Consequently, it became impractical to disentangle the independent effects of thermal versus non-thermal efficacy. Second, several LHT agents discussed in this review were characterized by insufficient supporting evidence. As reported by the CINeMA assessment, the certainty of evidence was rated as low or very low, for the majority of these therapies. Third, in all comparisons, exercise therapies were consolidated into a single node. However, unaccounted factors, such as differences in exercise training and the lack of supervision, may still have contributed to the heterogeneity observed in the exercise–control comparison (111, 112). Fourth, in the non-pharmacological or procedure-based interventions, identifying an appropriate control is more challenging, often leading to short-term or small-sample studies, and bias arising from inter-patient differences in thermal sensitivity was unavoidable. Fifth, over 70% of trials failed to implement a follow-up period exceeding 3 months, which limits the comprehensive understanding of the long-term effects of thermal interventions. Sixth, in the present study, although sensitivity analyses were conducted for trials utilizing the WOMAC subscale, separate NMAs for those using the VAS were limited by an insufficient number of trials, thereby compromising the comparability and interpretability in clinical practice. Thus, caution is warranted when interpreting the findings of the present study. The SUCRA value, which is highly sensitive to the structure of the NMA model and the degree of heterogeneity, possesses inherent limitations; hence, its results should also be interpreted with caution.

Conclusion

LHTs are effective non-pharmacological interventions for achieving immediate and short-term pain relief in adults with knee OA. Integrating LHTs with exercise therapy yields the most substantial and sustained clinical benefits compared with stand-alone treatments. Nevertheless, owing to the generally low quality of the included studies, well-designed, large-scale RCTs are warranted to rigorously validate their synergistic benefits.

ICMJE Statement of Interest

The authors declare that there is no conflict of interest that could be perceived as prejudicing the impartiality of the work reported.

Funding Statement

This work was supported by the National Natural Science Foundation of China (grant numbers: 82174406, 82274553, and 82374481) and the Shanghai Chronic Musculoskeletal Disorder Clinical Medical Research Center (20MC1920600). The funders had no role in the design, conduct, or decision to submit the manuscript for publication.

Author contribution statement

RZ designed the methodology; curated the data; performed formal analysis, investigation, and visualization; and wrote, reviewed, and edited the manuscript. LY was involved in conceptualization; performed formal analysis and visualization; and wrote, reviewed, and edited the manuscript. ZL conceived the study, designed the methodology, curated the data, and reviewed and edited the manuscript. JW was involved in data curation and reviewed and edited the manuscript. DJ performed investigation and reviewed and edited the manuscript. MY performed validation and reviewed and edited the manuscript. XW and ZS designed the methodology and reviewed and edited the manuscript. HZ acquired funding, administered the project, and reviewed and edited the manuscript. HG designed the methodology, administered the project, supervised the study, and reviewed and edited the manuscript. YC conceived the study, acquired funding, administered the project, supervised the study, and reviewed and edited the manuscript.

Data availability

All data relevant to the study are included in the article or are available as supplementary files.