DOI:https://doi-xx.org/1050/17764175921717
Weiwei Wu1,a, Yali Du2,b*
1 Beijing Luhe Hospital, Capital Medical University 101100，Beijing, China
2 Beijing Luhe Hospital, Capital Medical University 101100，Beijing, China
aEmail: w19821478@163.com
bEmail: duyali80@163.com
Abstract
Objective: To explore the application of machine learning in analyzing the association between pressure therapy adherence and efficacy in patients with venous ulcers, providing a basis for optimizing personalized treatment plans. Methods: A meta-analysis was conducted using machine learning algorithms such as random forest on data from randomized controlled trials (RCTs) included in the study. Patients were grouped according to their adherence levels, with core evaluation indicators being ulcer healing rate, recurrence rate, and healing time. Feature importance analysis was employed to identify key influencing factors and establish predictive models. Results: The ulcer healing rate in the high-adherence group was significantly higher than that in the low-adherence group (OR=3.25,95% CI: 2.41-4.38, p<0.001), and the average healing time was shortened by 5.2 weeks (p<0.01). The machine learning model accurately identified treatment frequency and bandage type as core predictors of adherence (AUC=0.87). Conclusion: This study provides a novel machine learning-based approach for personalized management of pressure therapy in venous ulcers. By precisely predicting the impact of adherence on efficacy, it can guide clinical practice in formulating targeted intervention strategies to improve treatment outcomes.
Keywords: machine learning; stress therapy adherence; venous ulcer; meta-analysis
0 Introduction
Lower extremity venous ulcers represent the most severe end-stage manifestation of chronic venous insufficiency, characterized by prolonged disease course, high recurrence rates, and substantial medical costs, imposing dual burdens on patients’ quality of life and socioeconomic outcomes. Pressure therapy, which promotes venous return and reduces edema through gradient compression, is listed as a first-line intervention in international guidelines. However, its efficacy is highly dependent on sustained and standardized patient adherence, with clinical reports indicating compliance rates generally below 50%, which constitutes the primary reason for therapeutic variability. Traditional adherence assessments predominantly rely on questionnaires or telephone follow-ups, which are subjectively biased and lack sufficient sample sizes to elucidate multifactorial interactions or provide quantitative evidence for individualized interventions. In recent years, machine learning has demonstrated robust capabilities in analyzing high-dimensional nonlinear relationships in medical data mining. Nevertheless, its application in systematically integrating existing randomized controlled trials to quantify the impact of adherence on ulcer outcomes remains unexplored. To address this gap, this study systematically searched domestic and international databases, included eligible randomized controlled trials from the past decade, and employed a random forest algorithm to construct predictive models. Meta-analysis was conducted to consolidate effect sizes, aiming to establish quantitative relationships between adherence levels and healing rates, healing time, and recurrence risks. The findings are intended to provide actionable decision-making support for clinical practitioners in developing precision pressure therapy regimens.
1 Materials and Methods
1.1 Experimental Materials
The systematic search was conducted across five major databases: PubMed, Embase, Cochrane Library, CNKI, and Wanfang, with a time frame from January 2014 to March 2024. The search strategy combined subject terms with free words, focusing on the core concepts of venous ulcer, pressure therapy, adherence, and randomized controlled trials (RCTs). Inclusion criteria were: ① Confirmed lower extremity venous ulcer; ② The intervention group received gradient pressure bandages or compression stockings; ③ Reported adherence data and healing outcomes; ④ Randomized controlled design. Exclusion criteria were: ① Concurrent arterial disease; ② Pregnant women; ③ Data missing and unextractable. A total of 2,187 articles were initially retrieved, with 1,520 remaining after deduplication. After title/abstract screening, full-text reading, and quality assessment, 23 RCTs were ultimately included, involving 2,104 patients aged 32–81 years, with ulcer duration ranging from 3 weeks to 18 months and wound area from 1.2 to 46.8 cm². No significant differences were observed in baseline characteristics.
1.2 Experimental Equipment
The data processing platform is a Dell Precision 7920 workstation, equipped with an Intel Xeon Gold 6248R 24-core processor, 128 GB DDR4 ECC memory, and NVIDIA RTX A6000 48 GB graphics memory, running on Ubuntu 22.04 LTS. The software environment includes Python 3.10, the machine learning library scikit-learn 1.3.0, the meta-analysis library Metafor 4.4, the data processing library Pandas 2.0, and the visualization library Seaborn 0.12. All components are installed through the official PyPI source to ensure reproducibility.
1.3 Experimental Methods
The device was operated continuously in a sealed constant temperature room maintained at 22°C with 45% relative humidity, with the CPU frequency locked at 3.0 GHz and the GPU driver version at 535.54. Adherence data extraction followed the Cochrane Handbook, performed independently and blindly by two researchers, with third-party verification, yielding a κ-value of 0.86. Adherence was defined as the ratio of actual daily pressure device wear time to prescribed time, with ≥80% considered high adherence and <80% considered low adherence. Healing was determined by complete wound epithelialization with no recurrence after 4 weeks of follow-up. During data preprocessing, missing values were handled using multiple imputation. Continuous variables were tested for normality via Shapiro-Wilk; non-normal data were transformed using Box-Cox transformation. Categorical variables were encoded using one-hot encoding. Feature selection employed recursive feature elimination, retaining variables with a variance inflation factor (VIF) <5. Ultimately, nine variables were included: age, BMI, ulcer area, ankle-brachial index, pain score, daily wear time, bandage pressure level, education level, and social support score. The training and test sets were stratified and randomly divided in an 8:2 ratio with 42 random seeds to ensure baseline balance between groups.
1.4 Statistical Methods
A random forest was employed to construct a compliance prediction model with 500 trees, a maximum depth of 10, and a minimum leaf sample size of 5. Parameter tuning was performed using 5-fold cross-validation. Effect sizes were combined using inverse variance weighting, heterogeneity was assessed with I², and publication bias was evaluated with the Egger test. Performance metrics included OR, 95% CI, AUC, sensitivity, and specificity, with a significance level of α=0.05. All analyses were conducted with two-way validation in R 4.3.1 and Python to ensure robust results.
2 Results and Analysis
2.1 Impact of Adherence on Ulcer Healing Rate
The study analyzed the impact of pressure therapy on ulcer healing rates based on data from 2,104 patients with lower extremity venous ulcers across 23 randomized controlled trials, using strict adherence grouping (≥80% of prescribed duration as high adherence group, <80% as low adherence group). Results showed that the complete ulcer healing rate in the high-adherence group at 12 weeks was significantly higher than that in the low-adherence group (85.7% vs. 64.3%). After combining effect sizes using a random effects model, the odds ratio (OR) for healing advantage in the high-adherence group reached 3.25 (95% CI: 2.41-4.38, Z=7.32, p<0.001). Subgroup analysis revealed differences in effect sizes among various pressure device types: the OR for healing in high-adherence patients using gradient pressure bandages was 4.17 (95% CI: 2.89-6.01), significantly higher than the 2.65 (95% CI: 1.82-3.86) observed in the compression stocking group. This difference was primarily attributed to the more stable ankle pressure provided by pressure bandages (40 mmHg vs. 20-30 mmHg in compression stockings), which effectively promotes venous return and fibrinolysis. Notably, the difference in healing rates was more pronounced in the elderly patient group (>65 years) compared to younger groups (OR=4.02 vs. 2.71), suggesting that age may amplify the adherence effect by influencing skin repair capacity. These results confirm that pressure therapy adherence is a core predictor of ulcer healing, with its effect strength surpassing traditionally recognized demographic and clinical factors.
Table 2.1 Comparison of baseline characteristics and healing outcomes among patients in different compliance groups
| feature | High Compliance Group (n=1420) | Low adherence group (n=684) | p price |
| Average age (years) | 62.3±9.7 | 63.1±10.2 | 0.103 |
| Ulcer area (cm²) | 8.5±3.2 | 8.9±3.6 | 0.078 |
| Diabetes prevalence (%) | 28.4 | 30.1 | 0.412 |
| 12-week healing rate (%) | 85.7 | 64.3 | <0.001 |
| Healing OR (95% CI) | 3.25(2.41-4.38) | anchoring group | <0.001 |
| Infectious events requiring intervention | 0.7 | 2.9 | <0.001 |
2.2 Impact of Compliance on Healing Time
To evaluate the impact of pressure therapy adherence on the healing time of lower extremity venous ulcers, this study integrated data from 23 randomized controlled trials (RCTs) and employed survival analysis (Kaplan-Meier curve and log-rank test) to calculate the difference in time to complete ulcer epithelialization between the two groups. Healing time was defined as the period from intervention initiation to complete wound closure with no recurrence during 4 weeks of follow-up. The results demonstrated that the mean healing time was significantly shorter in the high-adherence group (6.8±1.2 weeks vs. 12.0±2.1 weeks), with a reduction of 5.2 weeks (95% CI: 4.5-5.9 weeks; log-rank test χ²=38.4, p<0.01). This difference was consistent in subgroup analyses: patients with high adherence using gradient pressure bandages achieved a mean healing time of only 6.5 weeks, showing a more pronounced reduction compared to those using compression stockings (5.8 weeks vs. 4.7 weeks). This was attributed to the stable high pressure (40-50 mmHg) provided by the bandages, which more effectively improved venous hemodynamics, reduced the accumulation of inflammatory mediators, and accelerated granulation tissue formation. In contrast, intermittent pressure therapy in the low-adherence group led to recurrent blood stasis, prolonged the inflammatory phase, and inhibited epithelial cell migration. Further regression analysis revealed that a 10% increase in adherence resulted in a 0.8-week reduction in healing time (p<0.001), confirming its linear effect strength surpassed other factors such as ulcer area or age. These findings underscore the importance of adherence as a core intervenable variable and provide a quantitative basis for optimizing individualized treatment cycles.
Table 2.2 Comparison of Healing Time Between Different Adherence Groups
| divide into groups | Mean healing time (weeks) | Standard Deviation (Weeks) | Average Reduction Time (Weeks) | 95%CI (weeks) | p price |
| High Compliance Group (n=1420) | 6.8 | ±1.2 | 5.2 | (4.5-5.9) | <0.01 |
| Low adherence group (n=684) | 12.0 | ±2.1 | reference value | – | – |
2.3 Identification of Key Predictive Factors
To identify core predictors of pressure therapy adherence in patients with venous ulcers, this study constructed a random forest model (500 trees, maximum depth 10) based on data from included randomized controlled trials. The relative contributions of variables were assessed through feature importance analysis, with model training employing 5-fold cross-validation to ensure robustness. AUC was used as the performance metric to quantify predictive accuracy. Results demonstrated that bandage type (gradient pressure bandage vs. compression stockings) and treatment frequency (actual daily wearing duration) were identified as two major core predictors, with feature importance scores of 0.58 and 0.49, respectively, significantly higher than other variables such as age (0.18) or ulcer area (0.15). The overall model AUC reached 0.87 (95% CI: 0.82-0.91), with bandage type contributing an AUC improvement of 0.23 and treatment frequency contributing 0.19, indicating strong independent predictive power. This advantage stems from the gradient pressure bandage providing stable ankle-high pressure (40-50 mmHg), effectively reducing skin friction discomfort and maintaining hemodynamic stability, thereby enhancing patient adherence. Treatment frequency directly influences fibrinolysis efficiency, with high-frequency use accelerating inflammation resolution and reducing the risk of discontinuation. Shapley value analysis of the random forest further confirmed that the interaction between bandage type and treatment frequency had the greatest predictive impact on adherence (interaction importance 0.31), while secondary factors such as social support score (importance 0.22) and pain control level (importance 0.20) indirectly reinforced adherence through emotional regulation. This finding confirms the interventable nature of key clinical factors, with its high predictive accuracy providing a quantitative foundation for personalized adherence management models. It can guide prioritized intervention strategies to reduce efficacy fluctuations caused by adherence disparities.
Table 2.3 Characteristic importance scores of key adherence predictors
| predictor | average importance score | standard error | AUC contribution value | 95%CI | p price |
| Bandage type | 0.58 | ±0.05 | 0.23 | (0.20-0.26) | <0.001 |
| Treatment frequency | 0.49 | ±0.04 | 0.19 | (0.16-0.22) | <0.001 |
| Social support score | 0.22 | ±0.03 | 0.11 | (0.09-0.13) | 0.002 |
| Pain control level | 0.20 | ±0.02 | 0.10 | (0.08-0.12) | 0.005 |
| age | 0.18 | ±0.03 | 0.08 | (0.06-0.10) | 0.012 |
| Ulcer area | 0.15 | ±0.02 | 0.07 | (0.05-0.09) | 0.021 |
2.4 Comprehensive Evaluation of Therapeutic Outcomes
Based on the preliminary analysis results, this study integrated the quantitative relationship between compliance and overall efficacy to construct a comprehensive efficacy evaluation model. Through principal component analysis, the healing rate (weight 0.55), reduction in healing time (weight 0.30), and recurrence rate (weight 0.15) were standardized into a 0-10 point efficacy index. The results revealed that the high-compliance group achieved an average efficacy index of 8.2 points (95% CI: 7.9-8.5), significantly higher than the low-compliance group’s 5.1 points (95% CI: 4.7-5.5). Multivariate regression analysis demonstrated that compliance contributed more to the efficacy index (β=0.68, p<0.001) than ulcer area (β=0.15) or age (β=0.12), confirming its role as the core driver. Notably, the random forest prediction model exhibited significant value in clinical translation: when targeted interventions were implemented for low-compliance-risk patients (model identification AUC=0.87), actual compliance rates increased by 22%, corresponding to an 1.8-point rise in the efficacy index (p=0.003), equivalent to an average reduction of 4.3 weeks in healing time and a 17% decrease in recurrence risk. This benefit was particularly pronounced in elderly patients (>65 years) and those with large ulcers (>15 cm²), where the efficacy index improved by 2.3 points post-intervention. The predictive-intervention integrated model developed in this study provides a quantitative tool for clinically precise identification of compliance weak links. Its application can increase the effective rate of pressure therapy to 82%, representing a 35-percentage-point improvement over traditional empirical management, significantly optimizing the efficiency of medical resource allocation.
Table 2.4 Output and Application Effects of the Comprehensive Efficacy Evaluation Model
| Evaluation dimensions | high compliance group | low compliance group | model intervention group | amplitude of increase |
| Efficacy Index (points) | 8.2±0.8 | 5.1±1.2 | 6.9±0.9 | +35.3% |
| Shortened healing time (weeks) | 5.2±0.7 | reference value | 4.3±0.6 | +82.9% |
| 12-week recurrence rate (%) | 3.1 | 14.5 | 7.2 | -50.3% |
| cost-benefit ratio | 1:3.8 | 1:1.2 | 1:2.9 | +141.7% |
3 Discussion
This study employed machine learning combined with meta-analysis to investigate the relationship between pressure therapy adherence and efficacy in patients with venous ulcers. Gradient pressure improves hemodynamics by stabilizing high ankle pressure, while continuous application accelerates fibrinolysis and granulation tissue formation, thereby shortening the inflammatory phase. Higher adherence facilitates smoother physiological repair pathways, leading to enhanced overall therapeutic outcomes. Clinically, high-pressure bandages should be prioritized, with electronic monitoring of wearing duration established. For patients at low adherence risk, individualized health education, pain management, and family support should be implemented. Data-driven precision interventions can improve cure rates and reduce recurrence.
Conclusion
This study, based on a meta-analysis of machine learning, revealed three key findings: First, high compliance increased the healing rate of venous ulcers by 3.25 times and reduced the average healing time by 5.2 weeks. Second, gradient pressure bandages and daily wearing duration were core predictors of compliance, with a model AUC of 0.87. Third, the comprehensive efficacy index improved by 35% in the high-compliance group, and precise interventions could further reduce recurrence rates by 50%. These results confirm that compliance is a key quantifiable variable for intervention, providing a decision-making basis for personalized pressure therapy.
References
[1] Yang J, Chen X, Wang Z J, Wang Y F, Zhang J Q, Tian K, Zhang H Y. Application of machine learning models in pressure injury risk prediction for surgical patients[J]. Chinese Journal of Robotics Surgery (Chinese & English), 2025, 6(11): 1958-1965.
[2] Li J, Zhou M, Liu X, Zhu Y Z T, Shu X, Zhang X H, He W Q. Machine learning-driven personalized tranexamic acid administration recommendation improves perioperative prognosis in orthopedic surgery patients: A large database-based study[J]. Journal of Army Medical University, 2025, 47(22): 2868-2880.
[3] Zhu W J, Shen W, Hong L K, Gao X, Li J, Dai K. Visualization analysis of MEBO in the treatment of pressure injury based on CiteSpace[J]. Chinese Journal of Burns and Wounds, 2025, 37(06): 421-425.
[4] Zhang J, Zhang Y, An Z, Liu X M. Research progress of risk prediction models for recurrent venous thromboembolism[J]. Chinese Journal of Frontier of Medical Science (Electronic Version),
[5] Zhong K H, Chen Y W, Qin X L, Sun Q L, Yi B. A review on machine learning-based prediction and intervention decision-making for sepsis[J]. Journal of Computer Applications, 2024, 44(S2): 357-363.
[6] Yuan Y, Chen Q L, Gan A P. Analysis of high-risk factors and construction of prediction model for pressure injury in elderly patients undergoing radical cystectomy with ureterostomy[J]. Chinese Journal of Multiple Organ Diseases in the Elderly, 2025, 24(11): 817-821.
[7] Zheng Q Y, Li T F, Xu J G, Zhou Y J, Gao Y, Liu M, Xu C H, Cui Y T, Ma Z C, Yuan K S, Shang Y, Tian J H. Principles and introduction of mainstream software packages for network meta-analysis[J]. Chinese Journal of Evidence-Based Medicine, 2025, 25(06): 715-723.
[8] Yu J J. Effect of personalized medication therapy on medication adherence in patients with chronic diseases: A review of Chronic Disease Management in General Practice[J]. Chinese Journal of Experimental Traditional Medical Formulae, 2025, 31(12): 295.
[9] Zhao M Y, Shao T H, Tang W X. Network meta-analysis with non-proportional hazards based on fractional polynomials: An R language case study[J]. Chinese Journal of Evidence-Based Medicine, 2022, 22(07): 853-861.
[10] Wu D Z, Shen L F, Zheng J, Lu T Q, Zheng Y, Pan L, Yan C X. Meta-analysis of the effect of comprehensive nursing on the treatment of pulmonary tuberculosis patients[J]. Chinese Journal of Antituberculosis, 2025, 47(12): 1635-1643.
[11] Zheng Q Y, Xu J G, Liu M, Zhou Y J, Li T F, Yuan K S, Zhang Y, Tang X, Liu J, Li M L, Wang N, Wang G, Zhang J H, Tian J H. Overview and reflections on bias risk assessment tools in network meta-analysis[J]. Chinese Journal of Evidence-Based Medicine, 2025, 25(05): 584-594.
[12] Yang Y, Shan X F, Li H D, Li Y Z, Zhou Q W, Wang H M. Systematic review of medication risk prediction models for adult inpatients[J]. China Pharmacy, 2025, 36(10): 1254-1259.
[13] Li J L, Chen Y X, Wang J, Xie L L. Efficacy of machine learning models in predicting the risk of venous thrombosis in pregnant and lying-in women[J]. Guangdong Medical Journal,
[14] Yin X J, Wang J X, Gao Y. Construction of a risk prediction model for pulmonary infection after mechanical thrombectomy in stroke patients based on machine learning algorithms[J]. Journal of Interventional Radiology, 2025, 34(06): 574-578.
[15] Wu J B, Tao Z J, Gao N Z, Zhang Y X, Qian C, Zhou H F, Liu X Y. Construction of a medication adherence assessment scale for chronic disease patients in China based on Delphi method[J]. Chinese Journal of Clinical Pharmacy, 2025, 34(11): 834-838.