Diagnosis models to predict peripheral arterial disease: a systematic review and meta analysis

Abstract

Peripheral arterial disease (PAD) affects approximately 236.62 million individuals globally, exposing them to significantly increased risks of major limb events such as death and amputation. Concurrently, the number of diagnostic prediction models for PAD patients is steadily rising; however, these studies exhibit varying results, and their quality and applicability in clinical practice and future research remain unclear. To systematically assess the methodological quality of studies on PAD diagnostic prediction models. PubMed, Embase, Web of Science and Cochrane Database of Systematic Reviews were searched to identify studies which aiming to develop or validate a diagnostic prediction model of PAD. The retrieval time limit is from the establishment of the database to June 1, 2025. Two researchers independently screened and extracted data from eligible studies and evaluated the risk of bias using the Prediction Model Risk of Bias Assessment Tool (PROBAST). A total of 24 studies on PAD diagnostic prediction models were included, most of which exhibited high risk of bias, predominantly in the domains of study population and statistical analysis. The meta-analyzed Area Under the Receiver Operating Characteristic Curve (AUC) was 0.79 [0.74, 0.84], indicating favorable model performance. The reported number of predictor variables ranged from 2 to 20, with common predictors including age, gender, hypertension, diabetes, smoking, and BMI. This study demonstrates that PAD diagnostic prediction models exhibit good predictive performance, albeit accompanied by a high risk of bias and substantial heterogeneity across studies. Future research on modeling should emphasize comprehensive methodological enhancements in model design, construction, evaluation, and validation, with full disclosure of crucial model information. It should also utilize network computing for presenting model outcomes and conduct large-scale, multi-center external validation of existing models to promote their clinical application.

Trial registration: This study protocol has been registered with PROSPERO (registration number: CRD42024557144).

Background

Peripheral arterial disease (PAD) is a chronic progressive condition characterized by narrowing, obstruction, or dilation of peripheral arteries due to atherosclerosis and other factors, leading to ischemia, hypoxia, and functional impairments in limbs or organs, including lower extremity arterial occlusive disease (LEAD), renal artery stenosis (RAS), asymptomatic carotid stenosis(ACS), among other categories¹. Globally, PAD has become a significant public health issue, affecting approximately 236.62 million people, with an average annual prevalence rate of PAD at 10.7%, and increasing the risk of major adverse cardiovascular and cerebrovascular events by 3–6 times^2,3. Despite its prevalence, PAD remains underdiagnosed, with 40–60% of patients going undiagnosed in primary care settings^4,5. PAD often presents with asymptomatic or atypical symptoms, which can be mistaken for common signs of aging¹. The traditional gold standard for diagnosing PAD relies on imaging techniques to determine the location and severity of the lesions. However, this approach is invasive, costly, and not universally accessible⁶. The Ankle-Brachial Index (ABI) serves as an essential non-invasive parameter for evaluating PAD. It provides a hemodynamic assessment of lower extremity arterial perfusion. Key advantages include its non-invasiveness, rapidity, and low cost, rendering it highly suitable for large-scale population screening initiatives^1,6. However, limitations exist, notably its failure to accurately assess unilateral or focal lesions and its relatively low sensitivity for detecting early-stage or mild PAD. Timely and accurate diagnosis and treatment of PAD pose significant challenges in providing reliable healthcare services. One approach to facilitate early detection of PAD is through diagnostic prediction models. These models are highly valued in medical research for their ability to predict the probability of disease occurrence or outcomes based on individual clinical features and characteristics, utilizing non-invasive, cost-effective, and easily accessible markers with high sensitivity or specificity⁷. However, the utility of diagnostic prediction models for PAD is contingent upon their methodological quality. Models with high risk of bias not only fail to assess their actual benefits in clinical decision-making but may also lead to wastage of medical resources or misallocation. Therefore, this study aims to systematically review and evaluate PAD diagnostic prediction models, assessing their methodological quality to enhance their reliability and application value in clinical practice. By addressing these issues, this research seeks to contribute to the optimization and refinement of PAD diagnostic prediction models, facilitating their translation into effective clinical practice.

Methods

Inclusion and exclusion criteria

Inclusion Criteria: (1) Patients aged 18 years and older diagnosed with PAD. (2) Studies focused on clinical application-oriented modeling, including construction, validation, or (and) updating of prediction models, with at least 2 modeling variables. (3) Outcome measure of interest is PAD diagnosis. (4) Observational study designs such as cohort studies, case-control studies, and cross-sectional studies.

Exclusion Criteria: (1) Studies published as study protocols, conference papers, reviews, or abstracts. (2) Studies solely evaluating the methodological aspects of existing prediction models without presenting specific models. (3) Duplicate publications. (4) Literature inaccessible in full text.

Literature search strategy

Two researchers (QXY and XHR) independently conducted literature searches in PubMed, Embase, Web of Science, and the Cochrane Database to gather relevant studies on PAD diagnostic prediction models. The search covered articles published from the inception of each database to June 1, 2025. A combination of MeSH terms and free-text words was employed, with adaptations made according to database-specific features. The English search terms included: “peripheral arterial disease” OR “peripheral artery disease” OR “arteriosclerosis obliterans” OR “peripheral vascular disease” OR “intermittent claudication” OR “angina cruris” OR “lower extremity artery disease” OR “diagnosis” OR “diagnose” OR “diagnostic” OR “diagnosticate” OR “nomograms” OR “predict model” OR “predictive model” OR “prediction model” OR “model” OR “prediction” OR “risk factor” OR “risk score”. For PubMed, the specific search strategy is detailed in Fig. 1.

Fig. 1.

Literature search strategy expressions.

Literature screening and data extraction

Two researchers (QXY and XHR) conducted independent literature screening and data extraction, with cross-verification. Any discrepancies were resolved through discussion or consultation with a third person (SMH). The screening process involved initially reviewing titles and abstracts to exclude obviously irrelevant studies. Full-text articles were then reviewed to determine final inclusion based on relevance to the criteria outlined in the inclusion and exclusion standards. The data extraction was guided by the Checklist for critical Appraisal and data extraction for systematic Reviews of prediction Modelling Studies (the CHARMS checklist)⁸, and included the following components: basic information of studies(title, first author, publication year, study type, whether multicenter, inclusion and exclusion criteria of study participants, sample size, data sources, definition of the PAD diagnosis, and PAD diagnostic rate), prediction model characteristics(modeling type, modeling tools, modeling methods, validation methods, calibration methods, variable selection, handling of missing values, AUC values, sensitivity, specificity, final predictors and their quantities, and model presentation), risk of bias assessment key factors(factors related to bias risk assessment, including study population, predictors, outcome measures, and statistical analysis data).

Methodological quality assessment of included studies

The methodological quality of included studies was assessed using the Prediction Model Risk of Bias Assessment Tool (PROBAST)⁹, a standardized tool developed through a Delphi proces. PROBAST evaluates the methodological rigor of prediction model studies across four dimensions: study participants, predictors, outcome measures, and statistical analysis, through 20 key questions. Two researchers (QXY and XHR) independently applied PROBAST’s grading criteria, resolving any discrepancies through discussion or consultation with a third reviewer (SMH).

Statistical synthesis

For models reporting internally validated AUC values and their 95% confidence intervals, meta-analysis using Stata software (version 17.0) was performed with the “metan” command. Heterogeneity was assessed using the I² statistic and p-value to determine the variability across studies.

Results

Literature screening process and outcomes

Initially, a total of 1422 articles were identified. After systematic screening, 24 articles were selected for inclusion, comprising 17 studies on PAD^4,10–25, 1 on LEAD³, 2 on RAS^26,27, and 4 on ACS^28–31. Detailed information on the literature screening process and outcomes is illustrated in Fig. 2.

Fig. 2.

Literature screening flowchart.

Types and basic characteristics of included studies

The included studies were predominantly published between 2007 and 2024, with the majority originating from the America (58.33%) and China (29.17%). Retrospective studies accounted for a significant proportion (58.33%), and a substantial number were multicenter studies (54.17%). Sample sizes varied widely, ranging from 176 cases to 160.6 million cases, primarily sourced from databases (45.83%). Table 1 summarizes the basic characteristics of the included literature.

Table 1.

Table of the basic characteristics of the included literature.

First author and year	Country / region	The type of research	Whether multicenter	Inclusion criteria	Exclusion criteria	Sample size	Data sources	Definition of the PAD diagnosis	PAD diagnostic rate (%)
Sasikala P 2024	America	Retrospective study	No	① Cleveland heart disease dataset: comprising 14 features and 303 observations, the target variable is the result of invasive coronary angiography; ②statlog data set: consisting of 14 features and 270 observations, the target variable is the result of invasive coronary angiography	/	573 observations	Database	/	/
Xiong J 2021	China	Cross-sectional study	Yes	① Non-dialysis patients with CKD stage 1 to 5; ②adults (aged ≥ 18 years); ③complete medical records and follow-up data; ④informed consent	① Acute renal failure; ②previously or currently diagnosed with malignant tumor	1452	Previous studies	Atherosclerotic plaques are defined as focal structures invading the arterial lumen by at least 0.5 mm or 50% of the surrounding intima-media thickness value or with a thickness of > 1.5 mm measured from the media-adventitia interface to the intima-lumen interface	46.1%
Masoumi Shahrbabak S 2024	America	Retrospective study	No	Synthetic BP and PVR waveform signals related to abdominal aortic PAD generated through mathematical models	/	3,414,000 BP and PVR waveform signals	previous studies	The severity level of pad is defined by simulating the occlusion situation of the abdominal aorta through a mathematical model. when the occlusion degree is 0%, it indicates normal, and when it is 100%, it indicates complete occlusion	/
Ross EG 2016	America	Prospective study	No	Patients who underwent coronary angiography due to angina pectoris, dyspnea, or abnormal results of stress tests	Patients with incomplete or redundant data	1047	Hospital	ABI < 0.9	17.48%
McCarthy CP 2018	America	Prospective study	No	① Receive coronary and peripheral angiographic intervention; ②informed consent	/	354	Hospital	Obstructive pad: ≥ 50% stenosis in ≥ 1 peripheral blood vessel (carotid/subclavian artery, renal artery or lower extremity artery)	37.29%
Makdisse M 2007	Brazil	Prospective study	Yes	① Community-dwelling elderly people (≥ 75 years old); ②informed consent	① Situations affecting abi measurement: amputation, large ulcers, fractures, lower extremity pain, severe cognitive impairment, patient refusal; ②ABI > 1.40	176	Previous studies	Any lower extremity ABI ≤ 0.9	36.36%
Zhang Y 2016	America	Prospective study	Yes	① ABI assessment and lower extremity disease examination were limited to participants aged 40 and above; ②informed consent	① Bilateral amputation; ②obesity (> 400 pounds)	6982	Database	Any lower extremity ABI ≤ 0.9	Training set: 4.7%, validation set: 5.6%
Mansoor H 2018	America	Retrospective study	Yes	①≥ 40 years old; ②female	Lacking any covariate information or lacking ABIi value	3718 unweighted samples represent 150.6 million samples	Database	Any lower extremity ABI ≤ 0.9	13.68%
Wang F 2022	Britain	Retrospective study	Yes	① 40–69 years old; ②informed consent	①Failure of related relatives and X chromosome gender consistency checks; ②individuals with incompatible racial information and genetic information; ③individuals with more than 5% missing data; ④ambiguous SNPs, insertion-deletion SNPs and SNPs with more than 1% missing data or minor allele frequency less than 1%	487,320	Database	① Primary and secondary diagnoses and surgical codes for hospital events; ②use of PAD-specific drugs and self-reported PAD diagnoses and surgeries	1.18%
Gao JM 2022	China	Cross-sectional study	No	① > 80 years old; ②informed consent	Having rheumatic heart disease, heart valve disease, renal failure, stroke (within 3 months), acute and chronic infections, autoimmune diseases or tumors	539	Hospital	① Through lower extremity artery ultrasound examination, classification was carried out according to peak systolic velocity, velocity ratio, and degree of arterial stenosis, including normal (0% stenosis), 1–49% stenosis, 50–99% stenosis, and complete occlusion (100% stenosis); ②the hemodynamic important classification criterion for the diagnosis of “50–99% stenosis” requires that the peak systolic velocity at the lesion doubles compared to the more proximal segment (that is, greater than 200 cm/s, with evidence of turbulence)	46.75%
Weissler EH 2020	America	Retrospective study	No	At least one clinical visit has been conducted in this system and one or more PAD-related diagnostic codes have been generated: including peripheral vascular disease, atherosclerosis, diabetes with peripheral circulatory disorders, lower extremity ulcers, arterial thromboembolism and gangrene	/	2394	Electronic health record data system	ABI, previous history of revascularization or history of lower extremity amputation	32.58
Bali V 2016	America	Retrospective study	Yes	① Age ≥ 30 years old; ②newly diagnosed PAD patients should have no evidence of PAD within at least 12 months before the first diagnosis, and the date of the first diagnosis is defined as the index date; ③the control group was patients without PAD claims during the study period, and the date of their first claim formed the index date of the non-PAD group	① No continuous coverage within the 1-year baseline period (a 30-day gap is allowed); ②have any missing demographic information (gender and race)	139,610	database	①Use the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) diagnosis codes, ICD-9-CM procedure codes, and Current Procedural Terminology (CPT) codes to determine the PAD diagnosis; ②patients were identified as PAD cases if they had at least one hospitalization, outpatient visit, or professional claim	6.6%
Ghanzouri I 2022	America	Cross-sectional study	Yes	① Age ≥ 50 years old; ②at least 1 year of data; ③PAD patients: At least two separate ICD-9/ICD-10 or CPT codes and/or mentions of PAD in the notes and no exclusion codes; ④control group: No any PAD codes or text mentions in the health records	Data less than 1 year	20,031	Database	At least two separate ICD-9/ICD-10 or CPT codes and/or mentions of PAD in the notes	15.82%
Duval S, 2012	America	Prospective study	Yes	① Age ≥ 45 years old; ②patients with definite CVD, CAD, PAD, or those with three or more atherosclerotic risk factors.	① Have already participated in clinical trials; ②may have difficulty returning for follow-up or hospitalization	26,234	Database	In REACH: ①participants had intermittent claudication in the past or currently and ABI < 0.9 ②had received previous lower extremity arterial interventional therapy (such as leg angioplasty, stent implantation, atherectomy, peripheral arterial bypass surgery, or amputation) In FOS: ①ABI ≤ 0.9 ②presence of clinical PAD (intermittent claudication or lower extremity arterial revascularization)	① REACH Registry: 9.1%; ②Framingham Offspring Study: the definition of ABI ≤ 0.9: 8.0%; The definition of clinical PAD: 5.4%
Liang J 2022	China	Retrospective study	No	① Confirmed T2DM patients; ②age > 18 years old	① Type 1 diabetes or secondary diabetes; ②diabetes during pregnancy and lactation; ③seriously impaired consciousness or poor general condition; ④complicated with malignant tumors and heart, liver and kidney failure; ⑤incomplete clinical follow-up information	474	hospital	①Determined based on the current guidelines combined with auxiliary examinations; ②the examinations include ABI, limb vascular Doppler ultrasound, computed tomography angiography, vascular MRI or arteriography	Training set: 63.86%
Omiye JA 2024	America	Retrospective study	No	Informed consent	① ABI > 1.4 ; ②tibial vessels are not measurable	278	Database, electronic health record data system	There is a PAD diagnosis, ABI < 0.9, a history of vascular reconstruction surgery, and weakened foot pulse or wounds caused by arterial diseases in the medical record	18.71%
Ma Q 2015	China	Retrospective study	No	① Patients with ischemic heart disease; ②coronary angiography and renal artery angiography were performed during hospitalization; ③informed consent	/	355	Database	Renal artery angiography: Atherosclerotic arterial lesion, vascular diameter stenosis ≥ 60%	20%
Dong H 2018	China	Retrospective study	No	① Consecutive patients who received concurrent coronary angiography and renal artery angiography by the study doctor; ②informed consent	① Refusal of interventional coronary angiography or renal artery angiography; ② known or suspected non-atherosclerotic renovascular atherosclerotic syndrome; ③ regular renal replacement therapy, previous renal artery stent implantation, history of contrast-induced nephropathy; ④incomplete data	661	Hospital	Renal artery angiography: renal artery lumen stenosis of the main trunk or major branches ≥ 50%	Validation set: 72.76%
Ramos R 2011	Spain	Cross-sectional study	Yes	① 50–79 years old; ② participants without a history of cardiovascular diseases (angina pectoris, myocardial infarction, stroke or symptomatic peripheral artery disease	Participants with a history of cardiovascular diseases (angina pectoris, myocardial infarction, stroke or symptomatic PAD	7331	Database	ABI < 0.9	6.34%
de Weerd M 2014	Norway, Sweden, America, Germany	Prospective study	Yes	Age between 55 and 74 years old	Symptomatic patients	23,706	Previous studies	Measured by Doppler ultrasound combined with B-mode ultrasound imaging, moderate stenosis is defined as ≥ 50% stenosis, and severe stenosis is ≥ 70% stenosis	①≥50%: 2%; ②≥70%: 0.5%
Poorthuis MHF 2021	America, Britain, Netherlands	Retrospective study	yes	①Self-referred and self-paying individuals participating in the screening; ②ABI of 0.9 or lower	① Individuals who did not undergo ABI measurement or ABI determination was impossible (arteries could not be compressed); ②ABI was > 0.9 on both sides; ③individuals who did not undergo duplex ultrasound imaging or had inconsistent data	119,053	Vascular screening clinic, previous studies	① PAD is defined as ABI ≤ 0.9 on either side; ② moderate stenosis: ≥ 50%, based on a PSV of 150 cm/s or higher on each side, or 0 cm/s in case of arterial occlusion; Severe stenosis: ≥ 70%; based on a PSV of 210 cm/s or higher on each side, or 0 cm/s in case of arterial occlusion	①Derivation cohort : ≥ 50%: 5.7%, ≥ 70%: 2.5% ②validation queue1: ≥50%: 9.1%, ≥ 70%: 4.8% ③validation queue2: ≥50%: 16.8%, ≥ 70%: 12.4%
Poorthuis MHF 2021	America Britain	Retrospective study	Yes	①Provide blood samples; ②undergo ABI assessment and carotid duplex ultrasound examination	/	596,469	Vascular screening clinic	①≥50% stenosis: based on a PSV of 125 cm/s or higher (on either side) or 0 cm/s (occluded artery); ②≥70% stenosis: based on a PSV of 230 cm/s or higher (on either side) or 0 cm/s (occluded artery)	①≥50% ACS : 1.9%;②≥70% ACS: 0.3%
Kuang Y 2024	China	Retrospective study	Yes	Diabetic patient	Cases with incomplete data	3000	Database	/	15.9%
Cao Y 2024	China	Retrospective study	No	①Age 18–80 years; ②clinical manifestations of PAD; ③diminished arterial pulses; ④lower-extremity artery lesions (iliac arteries and distal); ⑤ABI ≤ 0.9; ⑥doppler ultrasonography/CTA-confirmed stenosis/occlusion.	①Lower extremity trauma/surgery history; ②severe cardiovascular/cerebrovascular diseases; ③hepatic/renal insufficiency; ④thyroid disorders; ⑤severe diabetes complications; ⑥malignancies; ⑦severe infections; ⑧autoimmune diseases; ⑨incomplete clinical data.	202	Hospital	Combined clinical symptoms, ABI ≤ 0.9, and imaging (doppler ultrasonography or CTA)	55.45%

① CKD: Chronic Kidney Disease; ② BP: Blood Pressure; ③ PVR: Pulse Volume Recording; ④ ABI: Ankle-Brachial Index; ⑤ SNPs: Single Nucleotide Polymorphisms; ⑥ CVD: Cerebrovascular Disease; ⑦ CAD: Coronary Artery Disease; ⑧ REACH: REduction of Atherothrombosis for Continued Health; ⑨ FOS: Framingham Offspring Study; ⑩ T2DM: Type 2 Diabetes Mellitus; ⑪ PSV: Peak Systolic Velocity;⑫CTA: Computed Tomography Angiography.

The majority of studies focused on model development and internal validation (75.00%), often employing R software for model construction (45.83%). Traditional modeling methods (75.00%) and machine learning (20.83%) were the predominant modeling methods, and one study jointly employed the two methods. Common specific modeling methods include multivariate logistic regression, random forest, and deep learning, among others. However, few studies adequately reported methods for handling missing data (30.00%). Approaches to addressing missing values included explicitly excluding data with missing values in exclusion criteria, multiple interpolation, and using the Keras framework, among others. Models typically incorporated between 2 and 20 predictive variables, commonly including age, gender, hypertension, diabetes, smoking, and BMI. Reported AUC values ranged from 0.58 to 0.96, with a significant majority reporting AUC values > 0.7 (79.17%). Score-based presentations were the most frequently used method for model display (50.00%). Table 2 provides an overview of the basic characteristics of the included study models.

Table 2.

Table of basic features incorporating the literature prediction model.

First author and year	Types of models	Modeling tools	Modeling method	Validation method	Calibration method	Variable screening method	Missing value processing	AUC (95%CI)	Sensitivity	Specificity	Final predictors and quantities	Model presentation
Sasikala P 2024	Modeling and internal validation	Jupyter Notebook Platform	SGD, SVM, K-NN, EDT, XGBoost, LR	10-Fold Cross-Validation	/	Hyper parameter optimization	/	/	Clevelan: 99.33%; Statlog: 98.09%	Cleveland: 99.12%; Statlog: 98.09%	/	/
Xiong J 2021	Modeling and internal validation	SPSS(25.0)	Logistic regression model	/		Binary logistic regression	/	Training set: 0.764(0.733–0.794); internal validation set: 0.808(0.765–0.852)	/	/	8:males, age, hypertension, diabetes, CKD stages, calcium, platelet, and albumin	Nomogram
Masoumi Shahrbabak S 2024	Modeling and internal validation	/	DL	/	/	/	/	PVR: ≥0.89 BP: :≥0.96	/	/	PVR: brachial and tibial artery PVR waveform; ABP: blood pressure waveform signal of brachial and tibial arteries	Score
Ross EG 2016	Modeling and internal validation	R(3.2.1)	The stepwise logistic regression models	10-fold cross-validation	HL test	Univariate analysis, forward stepwise logistic regression	/	The stepwise linear regression model : 0.76(0.68–0.85); penalized linear regression mode: 0.87 ( 0.81–0.92)	/	/	20:cumulative pack years; weakness in 1 or both legs: much; claudication present; ace inhibitors; unable to walk 50 ft; unable to walk 600 ft; weakness in 1 or both legs: very much; history of coronary artery disease; on temporary medical leave; history of cerebrovascular disease; current smoker; non-statin cholesterol agent; some walking impairment symptoms; calf/buttock pain, aching, or cramps: some; age; walking impairment symptoms: slight; calf/buttock pain, aching, or cramps: slight; unable to walk 1 block quickly; live alone; number of days engaged in rigorous activity	/
McCarthy CP 2018	Modeling and internal validation	R(3.4)	Machine learning (no specific method is reported)	Monte Carlo cross-validation	/	Least-angle regression, LASSO	/	Training set: 0.85;internal validation set: 0.84	65%	88%	7:history of hypertension, midkine, kidney injury molecule-1,interleukin-23, follicle-stimulating hormone, angiopoietin-1, and eotaxin-1)	math formula
Makdisse M 2007	Modeling and internal validation	SAS(8.2)	Multivariate logistic regression model.	/	HL test	Univariate logistic regression, multivariate logistic regression model, backward stepwise elimination method, forward stepwise variable	/	Internal validation set:0.85	85.90%	71.40%	4: abnormal pedal pulses, hypertension, cigarette smoking, and leg pain/discomfort in either leg on walking	score
Zhang Y 2016	Modeling and internal and external validation	Stata/MP(13.0)	Multiple logistic regression model	the bootstrap method, time verification	HL test	Likelihood ratio test	/	Training set: 0.82(0.82,0.83; external validation set: 0.76(0.72,0.79)	81.1%	72.2%	6: age, ethnicity, gender, pulse pressure, TC / HDL, and smoking status	math formula
Mansoor H 2018	Modeling and internal validation	SAS(9.4)	Multivariate logistic regression model	/	HL test, calibration curves	Backward stepwise elimination method	/	Training set: 0.74;internal validation set:0.73	/	/	7: age, BMI, hypertension, diabetes mellitus, smoking, non-oral contraceptive pill usage	score
Wang F 2022	Modeling and internal validation	R(3.6.3)	Generalized linear model	10-fold cross-validation	/	/	/	Internal validation set: 0.76(0.75–0.77)	/	/	10: age, gender, race / ethnicity, diabetes mellitus, BMI, hypertension, smoking status, CAD, CVD, and CHF	score
Gao JM 2022	Modeling and internal and external validation	/	LR, RF	/	/	/	/	Training set: ABI:0.880; LR: 0.866; RF: 0.951; internal validation set(based on the first seven features): 0.953	ABI:85.1%; IR: 81.5%; RF: 89.3%; based on the first seven features: 90.7%; RF external training set: 90.7%	ABI: 84.5%; LR: 83.8%; RF: 91.6%; based on the first seven features: 90.4%; RF external training se: 90.4%	7:ABI, creatinine, fasting blood glucose, age, history of CHD, diabetes mellitus and hypertension	/
Weissler EH 2020	Modeling and internal validation	SAS(/)	Binary logistic regression model	/	/	LASSO	/	Training set: 0.8618(0.8427–0.8810); internal validation set: 0.8618( 0.8352–0.8884)	75.30%	81.70%	15:ten separate ICD-9 / 10 CM codes, 1 combined ICD-9 + 10 CM code, and specialist visits with PAD related, prior revascularization, PAD-related diagnostic imaging and markers of 2 PAD-related visits during the study period	/
Bali V 2016	Modeling and internal and external validation	SAS(9.4)	Multivariate logistic regression model	/	calibration curves	Stepwise forward and backward elimination methods, multivariate logistic regression analysis	/	Internal validation set: 0.78; external validation set: 0.68	Internal validation set of 50% of the study cohort: 83.2%; internal validation set of 25% of the study cohort: 83.6%; external validation set: 98.0%	Internal validation set of 50% of the study cohort: 59.9%; internal validation set of 25% of the study cohort: 59.7%; external validation set: 8.6%	10:age, gender, diabetes mellitus, hypertension, dyslipidemia, chronic renal insufficiency, congestive heart failure, transient ischemic attack, angina pectoris and acute disease (acute myocardial infarction and stroke)	Screening indicators
Ghanzouri I 2022	Modeling and internal validation	R(3.6.3), Python(3.7.10)	Logistic regression model, RF, DL	Five-fold cross-validation	Calibration curves	/	LR: removes directly; RF: maintains the sparse matrix without entering missing values; DL: introduces the recent dimension and uses the Keras framework	Internal validation set: traditional logistic regression model : 0.81; RF:0.91; DL:0.96	Traditional logistic regression mode:0.75; RF:0.83; DL:0.95	Traditional logistic regression mode:0.73; RF:0.81; DL:0.88	Traditional logistic regression mode: 10: hypertension, hyperlipidemia, diabetes, CAD, CVD, CHF, BMI, race, age, sex RF: EHR data for all individuals (ICD-9 / 10, CPT, laboratory, drug and observation concept code); DL: 6: Age, race, sex, total number, features, clinical visit data, time proximity of diagnostic or laboratory values	Score
Duval S, 2012	Modeling and internal and external validation	SAS(9.1.3)	Multivariable stepwise logistic regression model	10-fold cross-validation	HL test	Multivariable stepwise logistic regression	/	Training set: 0.614; internal validation set: 0.603(0.602–0.605); external validation set: ABI: 0.64; FOS: 0.63	/	/	10:age, sex, smoking, diabetes mellitus, BMI, hypertension stage, and history of heart failure, CAD, CHF and CVD	Score, nomogram
Liang J 2022	Modeling and internal validation	R(4.4.1)	Multivariate logistic regression model	Bootstrap	HL test, calibration curves	Univariate logistic regression analysis, and multivariate logistic regression analysis	/	Training set: 0.765(0.711–0.819); internal validation set: 0.716(0.619–0.813)	/	/	3:disease course, blood urea nitrogen, and hemoglobin	Nomogram
Omiye JA 2024	Modeling and internal validation	R(3.6.3),Python(3.7.10)	Logistic regression model	/	Brier score	/	/	Internal validation set: clinical variables only: 0.902(0.846–0.957); clinical variables containing the PRS: 0.909(0.856–0.961)	/	/	12:CVD, CAD, CHF, CVD, CAD, CHF, hypertension, diabetes mellitus, hyperlipidemia, BMI, smoking status, age, gender, race, and PRS	Score
Ma Q 2015	Modeling and internal validation	SPSS(13.0)	Multivariate logistic regression model	/	HL test	Multivariate logistic regression	/	Internal validation set: predicted RAS was 0.808 and two-sided RAS was 0.762; external validation set: predicted RAS was 0.721 and two-sided RAS was 0.827	When the cut-off value was 35.0: ARAS 82.4% and two-sided ARAS 78.9%; external validation set: 85.0% ARAS and two-sided ARAS 85.7%	When the cut-off value was 35.0: RAS 51.0% and two-sided RAS 47.1%; external validation set: 30.5% RAS and two-sided RAS 17.5%	4:age, hypertension, stroke / intermittent claudication, and serum creatinine	Score
Dong H 2018	Modeling and internal and external validation	SAS(9.4), R(3.2.5)	Multivariate logistic regression model	bootstrap	HL test, calibration plots	Univariate analysis, collinear diagnosis	/	Internal validation set: 0.754(0.704,0.804) external validation set: 0.772(0.700,0.844)	97.00%	30.70%	4:hypertension, estimated glomerular filtration rate , early to late transmitral flow velocity ratio, low-density lipoprotein cholesterol	Nomogram
Ramos R 2011	Modeling and internal validation	R(2.0)	Multivariate logistic regression model	/	HL test	Univariate analysis, and a multivariate logistic regression analysis	/	Training set: 0.76(0.72–0.79); internal validation set: 0.76(0.73–0.79)	/	/	5:age, gender, smoking, pulse pressure, and diabetes mellitus	Score
de Weerd M 2014	Modeling and internal validation	/	Multivariate logistic regression model	bootstrap	HL test	Multivariate logistic regression	Single regression technology	Internal validation set ≥ 50%: 0.82(0.80–0.84); internal validation set ≥70%: 0.87(0.85–0.90)	/		8:age, gender, history of vascular disease, systolic, diastolic, TC / HDL ratio, diabetes mellitus and current smoking	Score, math formula
Poorthuis MHF 2021	Modeling and internal and external validation	STATA(15.1), R(3.5.1)	Multivariate logistic regression model	Bootstrap	Calibration plots	Stepwise forward elimination methods	Multiple interpolation	Internal validation set:≥50%: 0.71(0.71–0.72); ≥70%: 0.73(0.72–0.73); the first cohort external validation set:≥50%: 0.70(0.68–0.73);≥70%: 0.72(0.69–0.74); the second cohort for external validation:≥50%: 0.67(0.63–0.70); ≥70%: 0.67(0.64–0.70)	/	/	7:age, gender, smoking status, history of hypercholesterolemia, stroke / transient ischemic attack, CAD, and systolic blood pressure	Score, math formula
Poorthuis MHF 2021	Modeling and internal validation	STATA(15.1),R(3.5.1)	Multivariate logistic regression model	Bootstrap	Calibration plots	Multivariate logistic regression	Multiple interpolation	Internal validation set:≥50%: 0.78( 0.77–0.78); ACS ≥ 70%: 0.82(0.81–0.82)	When screening the 10% highest risk participants, 50% ACS was 40.9% and 70% ACS was 51.0%; when screening the 20% highest risk participants, the 50% ACS was 58.4% and the 70% ACS was 67.0%	When screening the 10% highest risk participants, 50% ACS was 90.6%; 70% ACS was 90.1%; when screening the 20% highest risk participants, 50% ACS was 80.7% and 70% ACS was 79.8%	9:age, gender, current smoking, diabetes mellitus, previous stroke / transient ischemic attack, CAD, PAD, blood pressure and blood lipids	Score, math formula
Kuang Y 2024	Modeling and internal validation	Python	LR, decision tree, RF, K-NN, and neural network algorithms	LR, decision tree, RF, K-NN, and neural network algorithms	/	Univariate analysis, multivariate logistic regression	exclude	LR: 0.73, decision tree: 0.58; RF: 0.68,; K-NN: 0.58; neural network algorithms: 0.72	/	/	7: male gender, atherosclerosis, carotid artery stenosis, fatty liver disease, hematologic diseases, endocrine disorders, elevated glycosylated serum protein	/
Cao Y 2024	Modeling and internal validation	R(4.2.2), SPSS (27)	Multivariate logistic regression model	/	Calibration curve	Univariate analysis, stepwise forward	Exclude	0.942	/	/	7: irisin, age, diabetes, dyslipidemia, smoking, creatinine, neutrophil/lymphocyte ratio	Nomogram

① SGD: Stochastic Gradient Descent;② SVM: Support Vector Machine;③ K-NN: K-Nearest Neighbors;④ EDT: Enhanced Decision Tree;⑤ XGBoost: Extreme Gradient Boosting; ⑥ LR: Logistic Regression; ⑦ DL: Deep Learning; ⑧ PLR: Penalized Linear Regression;⑨ FR: Random Forest; ⑩ ACEI: Angiotensin-Converting Enzyme Inhibitor;⑪ BMI: Body Mass Index;⑫ LASSO: Least Absolute Shrinkage and Selection Operator;⑬ T C/HDL: Total Cholesterol/High Density Lipoprotein;⑭ CHF: Congestive Heart Failure;⑮ PRS: polygenic risk scores;⑯ HL test: the Hosmer-Lemeshow Test:.

threshold.

Methodological quality assessment of included studies

In the assessment of participants, 19 studies were rated as high risk^{[3,4,10,11,16–27,29−31]}. This was primarily due to retrospective study designs and age restrictions that may have compromised the representativeness of the study populations. Regarding predictor variables, 16 studies did not clearly report whether assessments were blinded to outcome measures, resulting in an unclear bias risk^{[3,4,10,11,16,17,19–21,23−27,30,31]}. One study was rated unclear in bias risk due to unspecified predictive factors¹⁰. In terms of outcome assessment, five studies were deemed high risk for relying solely on ABI values for definitions^{12,14–16,25}, and two study did not specify PAD diagnostic criteria, leading to an unclear bias risk^3,10. Regarding statistical analysis, only three studies were categorized as low risk^25,27,30, while 12 studies were rated high risk due to issues such as inadequate sample sizes, discrepancies in reported multivariate analysis results compared to allocated predictive variables, and inadequate handling of missing data^{3,4,10,13,14,21–24,28,29,31}. 9 studies were classified as unclear due to incomplete reporting of predictive variables, their allocation, and missing data handling^{11,12,15–20,26}. Applicability assessment indicated only two study was unclear in predictive factors and outcome domains^3,10, with others deemed low risk. Bias risk and applicability assessment of included literature are summarized in Table 3.

Table 3.

The risk of literature bias and the suitability evaluation form.

Included in the literature	ROB				Applicability			Overall
	Participants	Predictors	Outcome	Analysis	Participants	Predictors	Outcome	ROB	Applicability
Sasikala P 2024	—	?	?	—	+	?	?	—	?
Xiong J 2021	+	+	+	—	+	+	+	—	+
Masoumi Shahrbabak S 2024	—	?	+	?	+	+	+	—	+
Ross EG 2016	+	+	—	?	+	+	+	—	+
McCarthy CP 2018	+	+	+	—	+	+	+	—	+
Makdisse M 2007	+	+	—	—	+	+	+	—	+
Zhang Y 2016	+	+	—	?	+	+	+	—	+
Mansoor H 2018	—	?	—	?	+	+	+	—	+
Wang F 2022	—	?	+	?	+	+	+	—	+
Gao JM 2022	—	+	+	?	+	+	+	—	+
Weissler EH 2020	—	?	+	?	+	+	+	—	+
Bali V 2016	—	?	+	?	+	+	+	—	+
Ghanzouri I 2022	—	?	+	—	+	+	+	—	+
Duval S, 2012	—	+	+	—	+	+	+	—	+
Liang J 2022	—	?	+	—	+	+	+	—	+
Omiye JA 2024	—	?	+	—	+	+	+	—	+
Ma Q 2015	—	?	+	?	+	+	+	—	+
Dong H 2018	—	?	+	+	+	+	+	—	+
Ramos R 2011	—	?	—	+	+	+	+	—	+
de Weerd M 2014	—	+	+	—	+	+	+	—	+
Poorthuis MHF 2021	—	?	+	+	+	+	+	—	+
Poorthuis MHF 2021	—	?	+	—	+	+	+	—	+
Kuang Y 2024	—	?	?	—	+	+	?	—	+
Cao Y 2024	—	?	+	—	+	+	+	—	+

“-” Is high risk, “+” is low risk, “?” is unclear.

Statistical integration

Due to insufficient reporting, meta-analysis was conducted on internal validation AUC values from 12 studies encompassing 17 predictive models^{12,17,19,22–25,27–31}, with subsequent analysis of heterogeneity. The pooled AUC estimate was 0.79 [0.74, 0.84], indicating significant heterogeneity among the predictive models. The forest plot displaying combined AUC estimates is presented in Fig. 3.

Fig. 3.

Forest plot of the AUC value Meat analysis. Ross EG 2016 (1)¹² refers to penalized linear regression mode, while Ross EG 2016 (2)¹² denotes a stepwise linear regression model. Omiye JA 2024 (1)²⁴ includes only clinical variables, whereas Omiye JA 2024 (2)²⁴ incorporates clinical variables along with PRS. de Weerd M 2014 (1)²⁹ indicates ≥ 50% threshold, whereas de Weerd M 2014 (2)²⁹ uses a ≥ 70% threshold. Poorthuis MHF 2021 (1)³⁰ employs a ≥ 50% threshold, and Poorthuis MHF 2021 (2)³⁰ utilizes a ≥ 70% threshold. Poorthuis MHF 2021 (1)³¹ employs a ≥ 50% threshold, and Poorthuis MHF 2021 (2)³¹ uses a ≥ 70%.

Discussion

Comparative analysis of PAD diagnostic prediction models

This study includes a total of 24 research studies, with most demonstrating moderate to good predictive performance (79.17% having an AUC > 0.7). The meta-analytic combined AUC value is 0.79 [0.74, 0.84], indicating strong discriminative ability of the prediction models in classifying samples relatively accurately. However, quality assessment reveals significant risk of bias across many studies, which may limit the clinical applicability of the respective prediction models. Each study exhibits specific strengths and weaknesses. For instance, Dong et al.²⁷conducted external validation and utilized vascular imaging indicators for PAD diagnosis, yet due to its retrospective nature, it carries high risks of selection and information biases, thus classified as high-risk. In contrast, Duval et al.²² derived their data from a multicenter prospective study but limited inclusion criteria to “age ≥ 45 years with clear atherosclerotic risk factors,” potentially compromising representativeness of real-world PAD patients and lacking transparency in handling missing data, also categorized as high-risk. The study by de Weerd et al.²⁹ employed a single regression technique for estimating missing values but did not report the modeling tool and imposed age restrictions in their inclusion criteria. Improvement opportunities exist across all models. Furthermore, in this study, DL yielded the highest AUC values, likely attributable to its ability to autonomously learn complex features from extensive data, achieve high-precision predictions with robust computational power, and adapt widely to various data types and tasks, thereby addressing data noise and uncovering hidden patterns^12,21. Conversely, multivariable stepwise logistic regression produced the lowest AUC values, potentially due to risks of overfitting, unstable outcomes, lack of theoretical basis in variable selection leading to potential information loss, and high sample size requirements²². It is noteworthy that while five studies reported AUC values > 0.7^{12,14–16,25}, they solely relied on ABI values for PAD diagnosis, introducing higher outcome bias risks. Given the complexity of PAD diagnosis, reliance solely on ABI values may overlook limitations such as susceptibility to arterial calcification affecting ABI readings, inadequate assessment of local lesions, and challenges in detecting early-stage lesions. In statistical analysis, inconsistencies between predictive variables and their allocation weights from five studies^{22,23,28,29,31}may indicate errors or biases in data processing, model construction, or analytical methods. Overall, despite demonstrating moderate to good performance, all models face considerable bias risks. Improvements are warranted in data sourcing, inclusion criteria, PAD diagnostic definitions, handling of missing values, allocation of predictive variable weights in final models, and reporting of multivariable analysis results. Additionally, significant heterogeneity among the predictive models in this study may stem from differences in study populations, predictive factors, and statistical methodologies.

Influence of age, sex, hypertension, diabetes, smoking, and BMI on predicting PAD diagnosis

Common predictive variables in PAD diagnostic prediction models include age, sex, hypertension, diabetes, smoking, and BMI. According to the latest guidelines from the American Heart Association, individuals aged ≥ 65 years face significantly increased risk of PAD, leading to a notable rise in disease incidence within this demographic³². With advancing age, alterations in arterial wall elastic proteins and collagen occur, resulting in reduced vascular compliance, increased lipid deposition, and gradual formation of atherosclerotic plaques^32,33. Regarding sex, males generally exhibit higher risk of PAD than females. Subgroup analysis by Xiong et al.²⁸ indicates distinct pathophysiological processes of atherosclerosis between sexes, potentially influenced by higher testosterone levels in males affecting endothelial function and promoting atherosclerosis, compounded by male-specific habits such as higher rates of smoking, alcohol consumption, poor dietary choices, engagement in physically demanding occupations, and exposure to environments predisposing vascular damage³⁴. Furthermore, hypertension and diabetes are pivotal risk factors for PAD. Prolonged hypertension subjects arterial vessel walls to sustained high pressure, reducing endothelial cell secretion of protective substances and increasing inflammatory factors, leading to endothelial damage³⁵. Endothelial injury promotes lipid deposition within vessel walls, forming atherosclerotic plaques, concurrently triggering proliferation and hypertrophy of vascular smooth muscle cells, exacerbating vascular narrowing³⁶. Diabetes contributes to accumulation of advanced glycation end products, impairs neurovascular function, damages endothelial cell function, and is often accompanied by insulin resistance, dyslipidemia, and heightened inflammatory responses, accelerating atherosclerosis³⁷. In addition, smoking accelerates the progression of atherosclerosis. Nicotine from tobacco stimulates sympathetic nervous system activity, inducing vascular spasm and constriction, while carbon monoxide reduces blood oxygen content, causing tissue hypoxia, promoting platelet aggregation, and thrombus formation^38,39. Studies demonstrate that smokers face several-fold higher risk of PAD compared to non-smokers, with risk increasing with greater smoking intensity and duration⁴⁰. Lastly, elevated BMI often correlates with obesity, which induces insulin resistance, reduces cellular sensitivity to insulin, disrupts glucose and lipid metabolism, and promotes lipid deposition within vessel walls, exacerbating endothelial damage due to increased inflammatory responses⁴¹.

Implications for future research construction and application

This study included a total of 24 studies, all conducted in economically developed countries such as the United States, United Kingdom, and China. There is a notable absence of research from low-income countries, and none of the studies included economic factors as predictive variables, potentially impeding the application of predictive models, given that socioeconomic status is a significant determinant of PAD occurrence⁴². In terms of model presentation, most studies utilized scoring and line charts, which are limited in interpretability, readability, and accessibility. Therefore, future research should consider employing web-based calculators that are convenient, widely accessible, cost-effective, and easily updated to enhance usability across a broader audience.

Additionally, the predominant modeling methods employed were logistic regression and machine learning. Logistic regression offers strengths such as high interpretability, computational efficiency, low data requirements, and model stability¹². Conversely, machine learning can handle complex nonlinear data relationships, adapt flexibly to various tasks, and manage imbalanced data, thereby achieving superior prediction accuracy. However, machine learning models are susceptible to issues like overfitting, opacity, data dependency, and poor interpretability. Future studies should address these limitations by increasing dataset sizes, conducting cross-validation, rigorously cleaning data, and utilizing techniques such as transfer learning and feature importance analysis to enhance research quality⁴³. Furthermore, despite 30.00% of studies in this research conducting external validation, this proportion remains relatively low. External validation is critical for assessing model generalizability, enhancing reliability and credibility, and facilitating model optimization and refinement. It offers a more realistic assessment of model applicability across diverse populations, settings, and medical contexts. Therefore, future research should prioritize multi-center external validation of existing models to further validate their generalizability, thereby providing robust evidence for clinical application and promoting standardization and advancement in medical research.

Lastly, the absence of comprehensive reporting of critical information poses challenges and introduces uncertainties in bias risk assessment for model evaluation. None of the studies in this research reported the time interval between the assessment of predictive variables and outcome determination, and most studies inadequately reported the allocation weights of predictive variables and the results of multivariable analyses in the final model. Hence, future research should strictly adhere to the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) statement to ensure standardized, comprehensive, and transparent reporting of research findings⁴⁴. In summary, future research should prioritize modeling studies in low-income countries, select appropriate modeling methods based on study characteristics such as linear relationships, data volume, and computational resources, present model results using web-based calculators, emphasize multi-center external validation of existing models, and adhere to the TRIPOD statement for standardized reporting, thereby enhancing the overall value of research endeavors.

Limitations of this study

This study searched only four databases. Despite their high authority and coverage, these databases may not encompass all relevant studies, potentially overlooking small-scale or grey literature databases from specific fields or regions. Additionally, due to differences in the transparency of included study reports and methodological approaches, our meta-analysis incorporated only 12 studies, which may limit the representativeness of the meta-analysis.

Conclusions

In summary, existing studies on predictive models for PAD diagnosis demonstrate acceptable model discrimination but generally exhibit poor methodological quality, with many studies posing high risks of bias. Factors contributing to bias risks include inappropriate data sources, variability in inclusion/exclusion criteria settings, reliance on single PAD diagnostic methods, inadequate handling of missing values, and incomplete reporting of information. Future modeling studies should comprehensively enhance methodological quality by adhering to PROBAST guidelines for designing, constructing, evaluating, and validating models. Furthermore, complete reporting of all critical model information as per the TRIPOD statement, presenting model results using web-based calculators, and emphasizing multicenter, large-sample external validation of existing models are recommended to facilitate the translational application of PAD diagnostic prediction models in clinical practice.

Author contributions

Conceptualization: QXY, XHR, SMH; Methodology: QXY, XHR, SP, CQ; Formal analysis and investigation: QXY, XHR, LXY, WD ; Writing - original draft preparation: QXY, LXY; Writing - review and editing: SP, WD, CQ, HXL; Funding acquisition: SMH; Resources: SMH; Supervision: HXL, SMH.

Funding

This study was supported by the Sichuan Science and Technology Program (Grant No. 2024NSFSC1605), Sichuan Provincial Key Laboratory of Nursing in 2023 (Grant No. HLKF2023 (Y)-3), Luzhou Science and Technology Program (Grant No. 2023RQN182).

Data availability

Data is provided within the manuscript or supplementary information files.

Declarations

Competing interests

The authors declare no competing interests.