Predicting radiocephalic arteriovenous fistula success with machine learning | npj Digital Medicine - Nature.com

1 year ago 69

Abstract

After instauration of a caller arteriovenous fistula (AVF), appraisal of readiness for usage is an important objective task. Accurate prediction of palmy usage is challenging, and augmentation of the carnal exam with ultrasound has go routine. Herein, we suggest a point-of-care instrumentality based connected instrumentality learning to heighten prediction of palmy unassisted radiocephalic arteriovenous fistula (AVF) use. Our investigation includes pooled patient-level information from 704 patients undergoing caller radiocephalic AVF creation, eligible for hemodialysis, and enrolled successful the 2014–2019 planetary multicenter PATENCY-1 oregon PATENCY-2 randomized controlled trials. The superior result being predicted is palmy unassisted AVF usage wrong 1-year, defined arsenic 2-needle cannulation for hemodialysis for ≥90 days without preceding intervention. Logistic, penalized logistic (lasso and elastic net), determination tree, random forest, and boosted histrion classification models were built with a training, tuning, and investigating paradigm utilizing a operation of baseline objective characteristics and 4–6 week ultrasound parameters. Performance appraisal includes receiver operating diagnostic curves, precision-recall curves, calibration plots, and determination curves. All modeling approaches but the determination histrion person akin favoritism show and comparable net-benefit (area nether the ROC curve 0.78–0.81, accuracy 69.1–73.6%). Model show is superior to Kidney Disease Outcome Quality Initiative and University of Alabama astatine Birmingham ultrasound threshold criteria. The lasso exemplary is presented arsenic the last exemplary owed to its parsimony, retaining lone 3 covariates: larger outflow vein diameter, higher travel volume, and lack of >50% luminal stenosis. A point-of-care online calculator is deployed to facilitate AVF appraisal successful the clinic.

Introduction

Functional vascular entree is indispensable for hundreds of thousands of patients successful the United States surviving with extremity signifier kidney illness (ESKD) and undergoing chronic intermittent hemodialysis1. Organized efforts to beforehand autogenous hemodialysis access, astir notably the National Kidney Foundation’s Kidney Disease Outcomes Quality Initiative (KDOQI), person resulted successful a displacement distant from prosthetic accesses and tunneled cardinal venous catheters (CVCs) toward arteriovenous fistulae (AVF)2,3. Despite efforts to beforehand the utilization of autogenous access, astir fractional of AVFs created are ne'er utilized successfully, and 80% of patients initiate hemodialysis with a CVC4,5,6,7. Underlying the sizeable saltation successful palmy AVF usage is uncertainty successful determining entree readiness for use. Newly created AVFs necessitate a play of maturation, wherever vas remodeling results successful a durable cannulation conception that tin beryllium utilized for hemodialysis8. Ideally, the maturation process takes spot implicit astir 6 weeks. Some AVFs volition necessitate further maturation clip oregon interventions anterior to usage owed to dilatory oregon maladaptive remodeling (e.g., intimal hyperplasia), portion others whitethorn beryllium utilized successfully without intervention. Significant acquisition is needed to find entree readiness with carnal exam, yet the availability of skilled providers is highly variable9,10. Uncertainty astir entree readiness whitethorn pb to prolonged dependence connected CVCs and either unnecessary oregon delayed interventions.

All members of the ESKD attraction team, including superior attraction doctors, nephrologists, surgeons, nurses, and technicians, should beryllium empowered to measure the wellness of hemodialysis accesses. To that end, rules for determining hemodialysis entree readiness for usage person been developed utilizing ultrasound arsenic a comparatively inexpensive, noninvasive, and elemental instrumentality to benchmark maturation progress. Existing rules are based connected static thresholds which spot patients into 2 categories: acceptable for usage and not yet acceptable for use. The commonly utilized existing thresholds see the anterior KDOQI criteria (≥600 mL/min travel volume, ≥6 mm diameter, and ≤6 mm heavy to skin) and the University of Alabama astatine Birmingham criteria (UAB; ≥ 500 mL/min travel measurement and ≥4 mm diameter)9,11. The existent enactment is motivated by designation that important accusation astir entree maturation is mislaid by dichotomizing the result of readiness for usage with static criteria. Additionally, the existing criteria were developed and validated successful smaller heterogenous cohorts with a number of forearm accesses, making their exertion to the radiocephalic AVF unclear5,9. Statistical models for prediction of AVF usage person been developed by the Hemodialysis Fistula Maturation (HFM) survey investigators, but the HFM observational cohort contained a number of forearm accesses5. Herein, we picture the improvement of a caller instrumentality for the prediction of palmy radiocephalic AVF usage which allows for a much nuanced objective mentation of entree readiness with improved prediction show erstwhile compared to some the UAB and KDOQI ultrasound threshold criteria.

Results

Summary statistics

The model-building cohort of those eligible for hemodialysis during the survey follow-up and with implicit 4–6 week ultrasound information comprised 591 patients (Fig. 1), 55% of whom were connected hemodialysis astatine the clip of AVF creation. The mean property was 57 (SD 13) years, 22% were female, and 65% were achromatic (Table 1). Radiocephalic AVFs were created astatine the wrist (75.3%), proximal forearm (22.3%), oregon anatomic snuffbox (2.4%). The mean intraoperative vein diameter was 3.37 mm (SD 0.82) and the mean artery diameter was 2.75 mm (SD 0.67). Patients with implicit 4–6 week ultrasound information successful the model-building cohort shared a akin covariate illustration with the wide cohort (n = 914, Table 1). Median follow-up successful the model-building cohort was 719 days (IQR 458–1068).

Fig. 1: Cohort enactment flowchart.
figure 1

Depicts travel from PATENCY-1 and PATENCY-2 randomized trials into prospective registry, eligibility for hemodialysis, receipt of postoperative ultrasound, and eventual introduction into predictive modeling cohort.

Table 1 Characteristics of each proceedings participants and prediction model-building cohort.

Ultrasound parameters

A full of 591 patients and 560 patients had implicit ultrasound information astatine 4–6 weeks and 12 weeks from the scale surgery, respectively (Fig. 1, Table 2). Among the model-building cohort, 277 patients (46.8%) achieved unassisted AVF usage wrong 1 year. Flow measurement was lowest successful those without AVF usage and highest successful those with unassisted AVF usage astatine some the 4–6 week ultrasound (mean quality 250 mL/min, 95% assurance interval [CI] 175–326, Fig. 2a) and 12 week ultrasound (mean quality 235 mL/min, 95% CI 136–334, Fig. 2b). Cephalic vein diameter was smallest successful those without AVF usage and largest successful those with unassisted AVF usage astatine some 4–6 weeks (mean quality 0.69 mm, 95% assurance interval [CI] 0.48–0.90, Fig. 2a) and 12 weeks (mean quality 0.95 mm, 95% CI 0.70–1.2, Fig. 2b). Flow measurement (mean quality 67.8 mL/min, 95% CI 39.5–96.1) and vein diameter (mean quality 0.43 mm, 95% CI 0.37–0.49) some accrued betwixt the 4–6 week and 12-week ultrasounds. Patients with palmy unassisted AVF usage were much apt to conscionable UAB (χ2[2] = 53.0, p < 0.001, Pearson’s Chi-squared test) and KDOQI (χ2[2] = 31.6, p < 0.001, Pearson’s Chi-squared test) criteria astatine their 4–6 week ultrasound. Among those with palmy unassisted AVF usage wrong 1-year, 73% and 52% did not conscionable KDOQI criteria astatine 4–6 weeks and 12 weeks, respectively.

Table 2 Successful radiocephalic arteriovenous fistula usage astatine 1-year by ultrasound parameters astatine 4–6 weeks and 12 weeks.
Fig. 2: Ultrasound parameters stratified by arteriovenous fistula (AVF) usage astatine 1-year.
figure 2

Points correspond idiosyncratic patients’ values. Data presented for 4 to 6 week a and 12 week b ultrasounds. Box plots picture median (center line), interquartile scope (box bounds, 25th and 75th percentiles), largest worth nary further than 1.5 times the interquartile scope (whiskers), and outlying values much than 1.5 times the interquartile scope plotted individually. Colors correspond AVF usage class (red: not utilized oregon indeterminate, orange: assisted use, blue: unassisted use). Shaded areas correspond to travel and diameter thresholds for UAB (≥500 mL/min and ≥4 mm, airy gray) and KDOQI (≥600 mL/min and ≥6 mm, acheronian gray) ultrasound criteria. Results of station hoc investigating comparing travel and diameter betwixt groups (Tukey’s test) are represented by asterisks (*p < 0.05).

Model Performance

Discrimination show of models based connected 4–6 week ultrasound measurements and baseline characteristics successful predicting 1-year unassisted AVF usage are shown successful Fig. 3 and Table 3. The show of UAB and KDOQI criteria approximations based connected travel and diameter thresholds were besides assessed. The Lasso exemplary (AUROC 0.794, AUPRC 0.719, accuracy 72.5%) performed astir arsenic good arsenic the elastic nett (AUROC 0.807, AUPRC 0.737, accuracy 71.3%) successful discriminating unassisted usage astatine 1 twelvemonth with a overmuch much parsimonious exemplary (Table 3). The favoritism show of each models was superior to that of KDOQI and UAB travel and diameter thresholds successful predicting 1-year unassisted AVF use, and the models demonstrated much balanced sensitivity and specificity. The Lasso exemplary had somewhat decreased calibration erstwhile compared to the elemental logistic regression model, but each models but the pruned histrion demonstrated acceptable calibration (Supplementary Fig. 1). Except for the pruned tree, each models had a higher net-benefit than either the no-information strategies (use each and usage none) oregon the existing static threshold strategies (UAB and KDOQI) crossed each tenable threshold probabilities (Supplementary Fig. 2). Additional exemplary details are included successful the supplementary materials (Supplementary Table 1, Supplementary Figs. 36).

Fig. 3: Model favoritism performance.
figure 3

Receiver operating diagnostic curves a, b and precision-recall curves c, d depicting exemplary show connected investigating information aft hyperparameter tuning and training. Panels are stratified by exemplary type, either logit a, c oregon histrion b, d for ocular clarity. Colors place the circumstantial exemplary (red: logistic regression, blue: lasso, orange: elastic net, green: random forest, purple: pruned classification tree, airy blue: boosted trees). Performance of University of Alabama astatine Birmingham (UAB; ≥ 500 mL/min and ≥4 mm) and Kidney Disease Outcomes Quality Initiative (KDOQI; ≥ 600 mL/min and ≥6 mm) ultrasound threshold criteria besides displayed arsenic labeled points.

Table 3 Comparison of exemplary show successful the hold-out investigating dataset.

Final prediction model

The Lasso logistic regression exemplary was favored arsenic the optimal prediction exemplary owed to its operation of show and parsimony. The last exemplary retained lone 3 covariates: larger outflow vein diameter successful US (per mm, likelihood ratio [OR] 1.95, 95% CI 1.48–2.60), higher travel measurement successful US (per 100 mL/min, OR 1.08, 95% CI 1.00–1.17), and lack of >50% luminal stenosis connected US (OR 2.74, 95% CI 1.65–4.60). No preoperative oregon intraoperative characteristics were retained by the Lasso. The value of the 3 ultrasound variables was shared crossed galore modeling approaches, with vein diameter and travel measurement contributing much to prediction show than each different variables successful the elastic net, pruned tree, random forest, and boosted histrion models (Supplementary Fig. 3). Predicted probabilities of palmy unassisted AVF usage by 1-year for an idiosyncratic diligent tin beryllium calculated utilizing the cross-table (Extended Data Supplementary Fig. 4), the nomogram (Supplementary Fig. 5), oregon the online calculator exertion (https://patrickheindel.shinyapps.io/predict-avf/).

Discussion

We applied instrumentality learning methods to prospectively collected information from rigorously conducted randomized objective trials to make a applicable instrumentality for estimating the probability of palmy unassisted radiocephalic AVF use. Our instrumentality incorporates accusation from standardized postoperative duplex ultrasounds and 23 baseline objective variables successful a cohort of 591 patients with recently created radiocephalic AVFs. The last prediction exemplary retained lone 3 predictors (AVF travel volume, vein diameter, and ≥ 50% stenosis measured astatine 4–6 weeks postoperatively utilizing duplex ultrasound) and exceeded the show of some the UAB and KDOQI ultrasound criteria. Additionally, our exemplary outputs predicted probabilities alternatively than strict people predictions (e.g., use/non-use), allowing for a much nuanced mentation of the output.

Prediction of palmy AVF usage is done routinely successful the session with carnal exam, but accuracy depends connected important accomplishment and experience, with adjacent the astir experienced clinicians achieving an accuracy of astir 80%9,10. Often, ultrasound is utilized to supplement carnal exam, and existing threshold criteria usher the appraisal of AVF readiness for use. Making the close appraisal has important implications for decisions regarding duration of CVC use, surgical oregon endovascular interventions, timing of hemodialysis initiation, and entree patency—all of which lend collectively to ESKD-related morbidity and mortality. Improving prediction of palmy unassisted AVF usage by supplementing objective assessments with point-of-care estimates based duplex ultrasonography should, therefore, elevate the prime of ESKD care. The improvement of the UAB and KDOQI criteria person unneurotic contributed substantially to ESKD diligent attraction and inspired the contiguous work. Recent advances successful prediction methodology, combined with the availability of high-quality granular information representing a challenging survey population, motivated our improvement of an updated attack to AVF assessment.

After tuning, fitting, and evaluating galore models, the Lasso was chosen arsenic our last model. The Lasso exemplary is appealing owed to its operation of simplicity and performance, with favoritism and calibration comparable to much analyzable methods similar random forest, and a net-benefit exceeding some UAB and KDOQI crossed a wide scope of thresholds. Discrimination refers to the quality of a exemplary to correctly differentiate betwixt cases and non-cases. AUROC is simply a metric that tin beryllium utilized to quantify the wide favoritism show of a model, and favoritism show implicit a scope of imaginable classification thresholds (Fig. 3a, b). A classification threshold is the predicted probability that, erstwhile exceeded, 1 would statement an reflection arsenic a case. When assessing metrics similar sensitivity, specificity, and affirmative predictive value, we chose a classification threshold of 0.5, a communal default for binary classification. Other classification thresholds could beryllium chosen by a clinician depending connected the objective request and penchant for prioritization of either sensitivity oregon specificity.

To licence valid predictions crossed a scope of classification thresholds, the estimated predicted probability indispensable bespeak the existent probability of the result successful the colonisation crossed each imaginable probabilities—this spot is called calibration. Calibration tin beryllium assessed visually by plotting deciles of predicted probabilities against the existent proportionality wrong that decile (Supplementary Fig. 1). A numeric appraisal of calibration tin beryllium made by calculating the slope and intercept of a exemplary regressing the result connected the predicted log-odds of the outcome, with the cleanable exemplary having slope = 1 and intercept = 0.

The prime of classification threshold is analogous to the prime betwixt utilizing either the KDOQI oregon UAB criteria—the KDOQI criteria mightiness beryllium utilized erstwhile favoring precocious specificity astatine the disbursal of sensitivity, portion the UAB criteria reflects a prioritization of sensitivity implicit specificity. Regardless, the model-based attack has a higher net-benefit crossed a wide scope of plausible classification thresholds than either KDOQI oregon UAB (Supplementary Fig. 2). In determination curve analysis, net-benefit represents a summary of fig of existent positives and mendacious positives and is utile successful summarizing some favoritism and calibration crossed a scope of thresholds12. The strategy with the highest nett payment crossed a plausible scope of threshold values volition beryllium the optimal prime for balancing existent positives and mendacious positives. Because applying this exemplary requires nary further information, cost, oregon investigating than what would beryllium indispensable for either the UAB oregon KDOQI strategies, the exemplary should beryllium preferred to these static criteria careless of the clinician’s threshold preference.

Our enactment confirms and extends the findings of anterior studies, possibly astir notably those of the HFM study, a multi-institution prospective observational cohort survey acrophobic with amended knowing AVF maturation13. Prior HFM enactment utilizing a backward elimination algorithm recovered that AVF travel volume, vein diameter, and extent from tegument were the astir important predictors of palmy AVF usage successful a mixed cohort including forearm (22.7%) and precocious limb (77.3%) AVFs5. A extremity of our survey was to spot if the summation of expanded baseline objective characteristics with much granular item would heighten predictive show successful a assortment of modeling approaches. Although the statistical methods employed by the HFM investigators disagree successful their details, our results look to replicate and corroborate the findings of the HFM study. Complex modeling strategies did not substantially amended the show of much parsimonious approaches, and ultimately, AVF travel measurement and diameter remained the astir important predictors of palmy AVF use.

This survey has immoderate cardinal strengths which should beryllium highlighted. The root information has precise debased missingness, precocious interior validity, and apt precise debased misclassification of predictors and outcomes owed to the prospective quality of information postulation for the purposes of research. Additionally, the illustration is homogenous with respect to entree configuration, with each participants undergoing caller instauration of a radiocephalic AVF. Restriction to lone radiocephalic AVF eliminates immoderate variability which whitethorn beryllium owed to inherent differences betwixt entree configurations - radiocephalic AVFs thin to beryllium distal, smaller, and with little travel volumes past brachiocephalic AVFs, for example. As noted above, our model’s show exceeds that of the static criteria without requiring further testing. The exemplary (PREDICT-AVF) is easy accessible and applicable for point-of-care applications done usage of the online calculator, cross-table, oregon nomogram illustration (https://patrickheindel.shinyapps.io/predict-avf/, Supplementary Figs. 4-5).

Our enactment indispensable beryllium interpreted with caution successful the discourse of the survey plan and inherent limitations. No underlying causal model guided our analysis, which was purely acrophobic with prediction. Readers should beryllium cautious to debar making causal interpretations oregon attributing excess meaning to the results of idiosyncratic components of immoderate prediction model. Additionally, definite simplifying assumptions were made to assistance successful the operation and mentation of the models which should beryllium kept successful mind. Competing events similar nonaccomplishment to follow-up, death, and renal transplantation were treated arsenic non-events successful this analysis. The accusation is that our exemplary predicts the probability of being observed to person palmy AVF use, alternatively than AVF usage itself. Unfortunately, the trials did not see the postulation of AVF extent accusation successful the ultrasound protocol, and though this parameter is portion of the anterior KDOQI “Rule of 6 s,” we had to approximate the accepted KDOQI criteria with lone travel measurement and diameter. Because each accesses successful this survey are radiocephalic, extent seems improbable to play arsenic important relation successful entree readiness for use, and lone 2.9% of patients successful either survey required a superficialization procedure. In addition, the PATENCY trials were conducted successful North America—caution should beryllium exercised erstwhile applying this exemplary to patients successful different settings, arsenic AV entree cannulation practices alteration importantly astir the world.

Finally, though the predicted probability of palmy AVF usage is of involvement to clinicians, the implications for however to usage this accusation to usher signifier are inactive unclear and warrant further investigation. For example, a clinician who sees a diligent with a predicted probability of palmy AVF usage of 30% whitethorn take to get further imaging, intervene with a surgical oregon endovascular process to assistance with maturation, wantonness the AVF, oregon simply hold and let much maturation time—we tin marque nary claims astir which of these strategies is optimal based connected the existent study. Any strategy prime needs to beryllium the effect of shared decision-making with the individualized ESKD life-plan successful mind2.

The contiguous survey contributes to ongoing enactment utilizing instrumentality learning techniques to amended ESKD care. Techniques applied successful our enactment tin beryllium readily expanded to different entree configurations and populations. Duplex ultrasound is an important non-invasive measurement of AVF maturation already successful regular use. Ultrasound measurements tin beryllium translated into interpretable estimates of unassisted usage occurrence done point-of-care tools developed with instrumentality learning.

Methods

Data source

We conducted a station hoc investigation of pooled patient-level information from the 2014–2019 planetary multicenter PATENCY-1 and PATENCY-2 signifier III randomized controlled trials (trial registration: ClinicalTrials.gov; NCT02110901, July 2014; and NCT02414841, August 2015). These trials prospectively tracked objective outcomes for up to 3 years pursuing caller radiocephalic AVF instauration astatine 31 and 39 centers, respectively, successful the United States and Canada. The superior trials’ elaborate methodology and results person been published previously14,15,16.

All precocious chronic kidney illness patients undergoing radiocephalic AVF instauration were eligible for enrollment successful the trials. Patients with a beingness expectancy of <6 months, progressive malignancy, oregon anterior attraction with the survey cause (vonapanitase, a recombinant quality elastase) were excluded from the trials. Ultimately, the proceedings cause vonapanitase was deemed to person constricted effect connected the applicable objective outcomes astatine 1 year, and further probe of the cause for this use-case was abandoned. Participants were followed prospectively for up to 3 years successful a pre-specified registry of objective outcomes. Enrollment began successful July 2014 and registry follow-up ended successful April 2019. Key information points collected during the proceedings and consequent registry follow-up included baseline comorbidities astatine clip of proceedings enrollment, anatomic and lawsuit premix characteristics, consequent surgical oregon endovascular interventions, and postoperative ultrasound measurements.

Routine duplex ultrasounds (US) were performed astatine 4 to 6 weeks and 12 weeks from AVF creation. Outflow-vein lumen diameter was measured doubly astatine 3 predetermined locations successful the forearm (3 cm proximal to the AVF anastomosis, mid-forearm, and instantly beneath the antecubital fossa) and averaged. Flow measurement was estimated from 3 abstracted measurements successful the aforesaid determination successful the cephalic vein 5 cm proximal to the AVF anastomosis. Stenosis was dichotomized arsenic beingness oregon lack of ≥50% luminal narrowing astatine immoderate constituent on the entirety of the access. Access extent was not assessed. All ultrasounds were interpreted by a blinded halfway laboratory (VasCore; Boston, MA). The methods were performed successful accordance with applicable guidelines and regulations, including waiver of informed consent, and approved by the Mass General Brigham quality probe committee Institutional Review Board for Use of antecedently collected proceedings information from PATENCY-1 and PATENCY-2 for station hoc analysis.

Prediction Models

We sought to physique upon and refine existing threshold-based ultrasound criteria for predicting AVF maturation and suitability for use. To beryllium included successful prediction modeling, patients needed to beryllium astatine hazard for AVF usage during the survey follow-up (e.g., connected hemodialysis) and person implicit 4- to 6-week ultrasound data. Any patients with pre-dialysis chronic kidney illness that did not advancement to requiring hemodialysis during the survey follow-up were excluded (Fig. 1).

Outcome

To amended interpretability and simplify exemplary building, the result for prediction modeling was dichotomized arsenic palmy unassisted AVF usage wrong 1-year, defined arsenic 2-needle cannulation for hemodialysis for ≥90 days without preceding intervention. Patients who did not successfully usage their AVF by 1 twelvemonth oregon anterior to a terminal lawsuit (death, transplant, entree abandonment, oregon nonaccomplishment to follow-up) were categorized arsenic not having palmy use. For patients with prevalent hemodialysis, the one-year clip model started connected the time of their surgery. For patients not yet receiving hemodialysis astatine the clip of AVF instauration and who did not commencement hemodialysis wrong one-year, palmy usage was defined arsenic 2-needle cannulation for each prescribed hemodialysis for a consecutive 90-day play starting wrong 6 weeks of hemodialysis initiation. Similar approaches person been implemented successful anterior analyses of AVF data5.

Covariate selection

Covariates were shared by each predictive modeling processes, and included age, sex, race, ethnicity, assemblage wide index, smoking status, aesculapian comorbidities, hemodialysis presumption astatine the clip of AVF creation, CVC history, CKD etiology, baseline vein and artery diameter measured successful the operating country aft induction of anesthesia, AVF location, anesthesia modality, anastomotic suture technique, statin use, antithrombotic use, and enrolling tract volume. Ultrasound information from the 4–6 week sojourn was chosen for predictive modeling due to the fact that of parallels with anterior enactment examining prediction of unassisted AVF use, objective relevance, and the complexity of including some 4–6 week and 12-week information unneurotic successful models. Ultrasound covariates included cephalic vein diameter, AVF travel volume, and the beingness oregon lack of ≥50% luminal stenosis. Analysis was restricted to patients with implicit 4–6 week ultrasound information arsenic described above. Covariate missingness was accounted for utilizing K-nearest neighbors imputation17.

Statistical analysis

In reporting descriptive statistics, categorical variables were summarized utilizing frequence with percentage. Continuous variables were reported arsenic mean with modular deviation erstwhile usually distributed, and median with interquartile scope otherwise. Unadjusted comparisons of ultrasound variables were made utilizing investigation of variance (ANOVA) followed by Tukey’s test. Paired information were compared utilizing paired t-tests. Categorical information were compared utilizing Pearson’s Chi-squared tests. A two-tailed alpha level of 0.05 was used. All investigation was performed utilizing R mentation 4.0.5 (https://cran.r-project.org/) and the packages tidyverse, tidymodels, glmnet, rpart, and ranger.

Modeling overview

To execute our extremity of gathering a predictive classification model, we explored respective modeling procedures each with its ain imaginable benefits and drawbacks. Modeling methods included accepted logistic regression, penalized logistic regression utilizing Lasso, classification, and regression histrion (CART) methods, and 2 ensemble classification methods: random wood and XGBoost. Each attack is chiseled with differing imaginable benefits and drawbacks; we sought to equilibrium exemplary complexity, flexibility, and show with interpretability and objective usefulness.

Multivariable logistic regression is utilized arsenic a “gold standard” successful classification problems. With respective covariates utilized for modeling, elemental logistic regression tin effect successful overfitting with bias successful coefficient estimation that leads to a driblet successful show erstwhile the exemplary is utilized connected outer data. To code this issue, penalized regression techniques usage coefficient shrinkage to trim out-of-sample bias; Lasso is simply a fashionable method owed to its quality to shrink coefficients to zero, acting arsenic an empiric adaptable enactment method and starring to simpler last models18. Notably, the bias-variance tradeoff volition ever beryllium a compromise and overfitting cannot beryllium eliminated, but penalization and cross-validation techniques described present tin mitigate overfitting (particularly successful smaller datasets).

The CART process is different accepted process for classification, with the cardinal payment of flexibly producing a clinically interpretable determination rule, but with a drawback of having perchance unstable show successful outer datasets adjacent with pruning methods19. To flooded this issue, histrion ensemble methods similar the random wood and XGBoost person been developed with wide adoption20,21. Random wood and XGboost are highly flexible and see interactions betwixt variables with comparatively debased bias. Random wood grows thousands of trees successful a akin mode to CART, but utilizing random samples of some variables and records which are past averaged implicit to execute a last exemplary (a method referred to arsenic bootstrap aggregating, oregon “bagging”). Similarly, XGBoost tin physique thousands of trees, but additionally uses the mistake from each histrion to reweight samples selected for each consequent histrion (referred to arsenic gradient boosting), theoretically preferencing variables with the astir predictive show and de-emphasizing meaningless variables. Variable value tin beryllium examined done a assortment of methods, but a deeper knowing of the relationships betwixt variables successful ensemble techniques is challenging and tin pb to skepticism from clinicians owed to little interpretability.

Modeling details

All predictive modeling methods were built with a training, hyperparameter tuning, and investigating paradigm utilizing a operation of baseline objective characteristics and 4–6 week US parameters described above. We performed a random 70/30 archetypal divided into grooming and investigating datasets anterior to exemplary building, diagnostics, oregon information cleaning. Continuous variables were preprocessed by centering (subtracting the average) and scaling (dividing by the modular deviation) their distributions anterior to exemplary fitting. A full of 5 missing values were imputed utilizing K-nearest-neighbors methodology (BMI, n = 1; intraoperative vein diameter, n = 2; intraoperative artery diameter, n = 2)17. Models were built utilizing the grooming dataset, and hyperparameters were tuned utilizing grid hunt methods with nested 10-fold cross-validation wrong the grooming dataset.

Our modeling attack started with elemental logistic regression including each covariates successful a main-effects model. Next, a lasso penalized logistic regression exemplary was acceptable to empirically prime covariates astir utile for prediction18. The regularization punishment was chosen to prime the astir parsimonious exemplary wrong 1 modular mistake of the regularization punishment with the minimum 10-fold cross-validated mean log-loss. Lasso was utilized for adaptable enactment for the refitting of a logistic regression model. Finally, an elastic nett exemplary was acceptable utilizing a regular grid hunt with 10 levels and nested 10-fold cross-validation to tune some the regularization punishment worth and the elastic nett mixing parameter22. Variable value was calculated arsenic the implicit worth of the scaled coefficients astatine the optimal regularization penalty.

A elemental classification histrion attack was besides pursued with the hopes of improving interpretability successful the lawsuit that a elemental and utile determination histrion could beryllium identified19. The histrion exemplary was pruned by optimizing the complexity parameter and histrion extent utilizing a regular grid hunt with 10 levels and nested 10-fold cross-validation. Variable value was calculated via the full Gini impurity simplification method.

A random wood classification exemplary was built with the extremity of expanding predictive show astatine the outgo of immoderate interpretability. Hyperparameters tuned included the fig of covariates for each attempted node divided and the minimum node size. Hyperparameters were tuned with a regular grid hunt with 10 levels and nested 10-fold cross-validation. All random wood models were built with 1,000 trees. Variable value was calculated via the Gini impurity simplification method20,23.

A boosted histrion exemplary was built utilizing the XGBoost method with a logistic nonaccomplishment function21. Tree depth, minimum node size, the learning rate, and the minimum nonaccomplishment simplification required to marque a further partition connected a leafage node were tuned utilizing nested 10-fold cross-validation and a maximum entropy grid hunt containing 100 hyperparameter configurations. Variable value was calculated via the accusation summation method.

After hyperparameter tuning, the last models were re-fit connected the full grooming dataset. The last exemplary show was assessed connected the prediction of the hold-out investigating dataset. A classification threshold of 0.5 was utilized for each models. Receiver operating diagnostic (ROC) curve plots, calibration plots, and determination curve plots were constructed for each modeling approach. Performance metrics of each modeling attack were calculated, including the country nether the ROC curve (AUROC), country nether the precision-recall curve (AUPRC), sensitivity, specificity, accuracy, and logistic calibration slope and intercept. The discriminative show of each exemplary was compared to the show of static threshold criteria approximating the UAB (flow measurement >500 mL/min and vein diameter >4 mm) and KDOQI (flow measurement >600 mL/min and vein diameter >6 mm) ultrasound criteria. Decision curves were plotted for each imaginable strategy for AVF usage prediction crossed a scope of threshold probabilities24.

Reporting summary

Further accusation connected probe plan is disposable successful the Nature Research Reporting Summary linked to this article.

Data availability

Limited deidentified information utilized for the analyses presented successful this enactment (training and investigating datasets) are disposable to qualified researchers connected request, delight email the corresponding writer Dr. Mohamad Hussain, MD, PhD astatine mhussain7@bwh.harvard.edu.

Code availability

Sample codification for each information processing and investigation presented successful this enactment are disposable connected request, delight email the corresponding writer Dr. Mohamad Hussain, MD, PhD astatine mhussain7@bwh.harvard.edu.

References

  1. Johansen, K. L. et al. US renal information strategy 2020 yearly information report: epidemiology of kidney illness successful the United States. Am. J. Kidney Dis. 77, A7–A8 (2021).

    Article  Google Scholar 

  2. Lok, C. E. et al. KDOQI objective signifier line for vascular access: 2019 update. Am. J. Kidney Dis. Off. J. Natl Kidney Found. 75, S1–S164 (2020).

    Article  Google Scholar 

  3. National Kidney Foundation–Dialysis Outcomes Quality Initiative. NKF-DOQI objective signifier guidelines for vascular access. Am. J. Kidney Dis. Off. J. Natl Kidney Found. 30, S150–S191 (1997).

    Google Scholar 

  4. Huber, T. S. et al. Arteriovenous fistula maturation, functional patency, and involution rates. Jama Surg. 156 (2021).

  5. Robbin, M. L. et al. Prediction of arteriovenous fistula objective maturation from postoperative ultrasound measurements: findings from the hemodialysis fistula maturation study. J. Am. Soc. Nephrol. 29, 2735–2744 (2018).

    Article  Google Scholar 

  6. Schinstock, C. A. et al. Outcomes of arteriovenous fistula instauration aft the Fistula First Initiative. Clin. J. Am. Soc. Nephrol. Cjasn 6, 1996–2002 (2011).

    Article  Google Scholar 

  7. Woodside, K. J. et al. Arteriovenous Fistula Maturation successful Prevalent Hemodialysis Patients successful the United States: A National Study. Am. J. Kidney Dis. Off. J. Natl Kidney Found. 71, 793–801 (2018).

    Article  Google Scholar 

  8. Donnelly, S. M. & Marticorena, R. M. When is simply a caller fistula mature? The emerging subject of fistula cannulation. Semin Nephrol. 32, 564–571 (2012).

    Article  Google Scholar 

  9. Robbin, M. L. et al. Hemodialysis arteriovenous fistula maturity: US evaluation. Radiology 225, 59–64 (2002).

    Article  Google Scholar 

  10. Ferring, M., Henderson, J. & Wilmink, T. Accuracy of aboriginal postoperative objective and ultrasound introspection of arteriovenous fistulae to foretell dialysis use. J. Vasc. Access 15, 291–297 (2013).

    Article  Google Scholar 

  11. Vascular Access Work Group. Clinical signifier guidelines for vascular access. Am. J. Kidney Dis. 48, S176–S247 (2006).

    Article  Google Scholar 

  12. Calster, B. V. & Vickers, A. J. Calibration of hazard prediction models: interaction connected decision-analytic performance. Med. Decis. Mak. Int J. Soc. Med. Decis. Mak. 35, 162–169 (2014).

    Article  Google Scholar 

  13. Dember, L. M. et al. Objectives and plan of the hemodialysis fistula maturation study. Am. J. Kidney Dis. Off. J. Natl Kidney Found. 63, 104–112 (2013).

    Article  Google Scholar 

  14. Bleyer, A. J. et al. A randomized proceedings of vonapanitase (PATENCY-1) to beforehand radiocephalic fistula patency and usage for hemodialysis. J. Vasc. Surg. 69, 507–515 (2019).

    Article  Google Scholar 

  15. Peden, E. K. et al. PATENCY-2 proceedings of vonapanitase to beforehand radiocephalic fistula usage for hemodialysis and secondary patency. J Vasc Access 1129729820985626 https://doi.org/10.1177/1129729820985626 (2021).

  16. Heindel, P. et al. Radiocephalic arteriovenous fistula patency and use: a station hoc investigation of multicenter randomized objective trials. Ann. Surg. Open 3, e199 (2022).

    Article  Google Scholar 

  17. Gower, J. C. A wide coefficient of similarity and immoderate of its properties. Biometrics 27, 857 (1971).

    Article  Google Scholar 

  18. Tibshirani, R. Regression shrinkage and enactment via the lasso. J. R. Stat. Soc. Ser. B Methodol. 58, 267–288 (1996).

    Google Scholar 

  19. Breiman, L., Friedman, J. H., Olshen, R. A. & Stone, and C. J. Classification And Regression Trees. (Wadsworth International Group, 1984).

  20. Breiman, L. Random Forests. Mach. Learn 45, 5–32 (2001).

    Article  Google Scholar 

  21. Chen, T. & Guestrin, C. XGBoost: A Scalable Tree Boosting System. Arxiv 785–794 https://doi.org/10.1145/2939672.2939785 (2016).

  22. Zou, H. & Hastie, T. Regularization and adaptable enactment via the elastic net. J. R. Stat. Soc. Ser. B Stat. Methodol. 67, 301–320 (2005).

    Article  Google Scholar 

  23. Wright, M. N., Ziegler, A. & König, I. R. Do small interactions get mislaid successful acheronian random forests? Bmc Bioinforma. 17, 145 (2016).

    Article  Google Scholar 

  24. Vickers, A. J. & Elkin, E. B. Decision curve analysis: a caller method for evaluating prediction models. Med Decis. Mak. 26, 565–574 (2006).

    Article  Google Scholar 

Download references

Acknowledgements

This enactment was funded successful portion by a NIH T32 HL007734 fellowship grant to Patrick Heindel, MD. Mohamad Hussain, MD, PhD, is funded by a Brigham and Women’s Hospital Heart and Vascular Center Junior Faculty Award.

Author information

Authors and Affiliations

  1. Division of Vascular and Endovascular Surgery, Department of Surgery, Brigham and Women’s Hospital, Harvard Medical School, Boston, MA, USA

    Patrick Heindel, Jessica D. Feliz, Michael Belkin, C. Keith Ozaki & Mohamad A. Hussain

  2. Center for Surgery and Public Health, Brigham and Women’s Hospital, Boston, MA, USA

    Patrick Heindel, Tanujit Dey, Jessica D. Feliz & Mohamad A. Hussain

  3. Division of Renal Medicine, Department of Medicine, Brigham and Women’s Hospital, Harvard Medical School, Boston, MA, USA

    Dirk M. Hentschel

  4. Division of Cardiovascular Medicine, Department of Medicine, Brigham and Women’s Hospital, Harvard Medical School, Boston, MA, USA

    Deepak L. Bhatt

  5. Division of Vascular Surgery and Li Ka Shing Knowledge Institute, St. Michael’s Hospital, University of Toronto, Toronto, ON, Canada

    Mohammed Al-Omran

  6. Department of Surgery, King Faisal Specialist Hospital and Research Center, Riyadh, Saudi Arabia

    Mohammed Al-Omran

Contributions

P.H. drafted the manuscript text. P.H., T.D., and M.A.H. performed statistical analyses. D.M.H., M.B., and C.K.O. substantially contributed to information acquisition and survey design. All authors (P.H., T.D., J.D.F., D.M.H., D.L.B., M.A.O., M.B., C.K.O., M.A.H.) contributed to the conception and mentation of the work, manuscript revision, last approval, and attest to the works’ accuracy and integrity.

Corresponding author

Correspondence to Mohamad A. Hussain.

Ethics declarations

Competing interests

The authors state nary Competing Financial Interests but the pursuing Competing Non-Financial Interests: Dr. Deepak L. Bhatt discloses the pursuing relationships - Advisory Board: AngioWave, Bayer, Boehringer Ingelheim, Cardax, CellProthera, Cereno Scientific, Elsevier Practice Update Cardiology, High Enroll, Janssen, Level Ex, Medscape Cardiology, Merck, MyoKardia, NirvaMed, Novo Nordisk, PhaseBio, PLx Pharma, Regado Biosciences, Stasys; Board of Directors: AngioWave (stock options), Boston VA Research Institute, DRS.LINQ (stock options), High Enroll (stock), Society of Cardiovascular Patient Care, TobeSoft; Chair: Inaugural Chair, American Heart Association Quality Oversight Committee; Data Monitoring Committees: Acesion Pharma, Assistance Publique-Hôpitaux de Paris, Baim Institute for Clinical Research (formerly Harvard Clinical Research Institute, for the PORTICO trial, funded by St. Jude Medical, present Abbott), Boston Scientific (Chair, PEITHO trial), Cleveland Clinic (including for the ExCEED trial, funded by Edwards), Contego Medical (Chair, PERFORMANCE 2), Duke Clinical Research Institute, Mayo Clinic, Mount Sinai School of Medicine (for the ENVISAGE trial, funded by Daiichi Sankyo; for the ABILITY-DM trial, funded by Concept Medical), Novartis, Population Health Research Institute; Rutgers University (for the NIH-funded MINT Trial); Honoraria: American College of Cardiology (Senior Associate Editor, Clinical Trials and News, ACC.org; Chair, ACC Accreditation Oversight Committee), Arnold and Porter instrumentality steadfast (work related to Sanofi/Bristol-Myers Squibb clopidogrel litigation), Baim Institute for Clinical Research (formerly Harvard Clinical Research Institute; RE-DUAL PCI objective proceedings steering committee funded by Boehringer Ingelheim; AEGIS-II enforcement committee funded by CSL Behring), Belvoir Publications (Editor successful Chief, Harvard Heart Letter), Canadian Medical and Surgical Knowledge Translation Research Group (clinical proceedings steering committees), Cowen and Company, Duke Clinical Research Institute (clinical proceedings steering committees, including for the PRONOUNCE trial, funded by Ferring Pharmaceuticals), HMP Global (Editor successful Chief, Journal of Invasive Cardiology), Journal of the American College of Cardiology (Guest Editor; Associate Editor), K2P (Co-Chair, interdisciplinary curriculum), Level Ex, Medtelligence/ReachMD (CME steering committees), MJH Life Sciences, Oakstone CME (Course Director, Comprehensive Review of Interventional Cardiology), Piper Sandler, Population Health Research Institute (for the COMPASS operations committee, publications committee, steering committee, and USA nationalist co-leader, funded by Bayer), Slack Publications (Chief Medical Editor, Cardiology Today’s Intervention), Society of Cardiovascular Patient Care (Secretary/Treasurer), WebMD (CME steering committees), Wiley (steering committee); Other: Clinical Cardiology (Deputy Editor), NCDR-ACTION Registry Steering Committee (Chair), VA CART Research and Publications Committee (Chair); Research Funding: Abbott, Acesion Pharma, Afimmune, Aker Biomarine, Amarin, Amgen, AstraZeneca, Bayer, Beren, Boehringer Ingelheim, Boston Scientific, Bristol-Myers Squibb, Cardax, CellProthera, Cereno Scientific, Chiesi, CSL Behring, Eisai, Ethicon, Faraday Pharmaceuticals, Ferring Pharmaceuticals, Forest Laboratories, Fractyl, Garmin, HLS Therapeutics, Idorsia, Ironwood, Ischemix, Janssen, Javelin, Lexicon, Lilly, Medtronic, Merck, Moderna, MyoKardia, NirvaMed, Novartis, Novo Nordisk, Owkin, Pfizer, PhaseBio, PLx Pharma, Recardio, Regeneron, Reid Hoffman Foundation, Roche, Sanofi, Stasys, Synaptic, The Medicines Company, 89Bio; Royalties: Elsevier (Editor, Braunwald’s Heart Disease); Site Co-Investigator: Abbott, Biotronik, Boston Scientific, CSI, Endotronix, St. Jude Medical (now Abbott), Philips, Svelte, Vascular Solutions; Trustee: American College of Cardiology; Unfunded Research: FlowCo, Takeda. Dr. C. Keith Ozaki discloses the pursuing relationships –Proteon Therapeutics Inc, technological advisory board; Humacyte, Inc, advisory board, consultant; Medtronic, consultant; Laminate Medical Technologies, consultant. Dr. Dirk M. Hentschel discloses the pursuing relationships – Bard BD, consultant; BluegrassVascular, consultant; Laminate Medical, consultant; Medtronic, consultant; Merit, consultant; Sanifit, consultant; Shifamed, consultant; Surmodics, consultant; VenoStent, consultant; Humacyte Inc, advisory board; Nephrodite, advisory board.

Additional information

Publisher’s note Springer Nature remains neutral with respect to jurisdictional claims successful published maps and organization affiliations.

Supplementary information

About this article

Verify currency and authenticity via CrossMark

Cite this article

Heindel, P., Dey, T., Feliz, J.D. et al. Predicting radiocephalic arteriovenous fistula occurrence with instrumentality learning. npj Digit. Med. 5, 160 (2022). https://doi.org/10.1038/s41746-022-00710-w

Download citation

  • Received: 20 June 2022

  • Accepted: 10 October 2022

  • Published: 25 October 2022

  • DOI: https://doi.org/10.1038/s41746-022-00710-w

Read Entire Article