Gaëlle Romain, PhD, MSc

Background The Vascular Quality Initiative (VQI)’s linked Medicare registry (VISION), derived CPT-based algorithms for their amputation outcomes in peripheral vascular intervention (PVI), supra-inguinal bypass (SUPRA) or infra-inguinal bypass (INFRA) modules. We examined the validity of the amputation algorithm against medical chart abstraction. Methods In VISION, major amputation was defined by CPT codes 27590, 27591, 27592, 27880, 27881, and 27882 in the PVI, SUPRA, and INFRA modules for procedures performed between 2010 and 2017 in patients ≥18 years. We compared the number of amputations at 1, 3, and 5 years following the index revascularization procedure in these modules against Yale New Haven Hospital (YNHH) and Dartmouth Hitchcock Medical Center (DHMC) YNHH-DHMC chart abstraction as the gold standard. Concordance was examined using Spearman's rank correlation (ρ) and Cohen's kappa (κ) statistic. Results Medical records of 1,189 PVI, 247 INFRA and 93 SUPRA procedures were reviewed. Concordance for the number of amputations was moderate to strong (ρ.63-.75, p<.001), as well as in major amputation (κ.56-.73, p<.001) between VISION and YNHH-DHMC chart abstraction (Table). Conclusion Concordance between the number of major amputations between VISION and YNHH-DHMC chart abstraction was relatively high, but not perfect. Further augmentation of the CPT-based algorithm may be required to improve the quality of major amputation data in VISION.

Introduction: Individualized prediction models for transfemoral carotid artery stenting (TF-CAS) health status outcomes have not been developed. We applied a machine learning algorithm to identify the most robust pre-procedural predictors of 30-day TF-CAS health status, and validated models across different health status measures to maximize generalizability. Methods: The 390-center SAPPHIRE registry enrolled high-risk patients undergoing TF-CAS from 2006-2014, containing two cohorts with pre-procedural and 30-day health status assessments: Cohort 1 (n=4,667) had EQ-5D-3L (Index and Visual Analog Scale [VAS]); Cohort 2 (n=3,594) had SF-36 (Mental [MCS] and Physical [PCS] Component Summary scores). A random forest algorithm ranked the importance of 50 pre-procedural variables (pre-procedural health status, modified Rankin and NIH stroke scores, demographics, comorbidities, and disease characteristics) for 30-day health status in each cohort. Selection thresholds were based on visual inspection of the importance plots to evaluate the information gain value of each variable considered. Overlapping variables across health status measures were retained to develop a multivariable linear regression model to predict 30-day health status in each cohort. Results: Importance plot (Figure) inspection yielded 23 variables, each of which had importance values >15% for ≥2 of the 4 health status measures. From these, 7 overlapping variables were ultimately retained in the linear regression analyses that predicted ~40% of the variance (R2=.39-.45) for 30-day TF-CAS health status, including pre-procedural health status, age, heart failure, history of stroke, Rankin stroke score, renal disease, and pulmonary disease. Conclusions: Seven easily assessed pre-procedural variables robustly predicted TF-CAS health status. This model can inform the risk-benefit analysis underlying the TF-CAS medical decision-making process for patients and clinicians.

Context: In the field of population-based cancer epidemiology, net survival (Sn) is modelled through the excess hazard (EH). For many cancer sites, a proportion of patients will not die from the studied cancer, representing the cured proportion (P). Cure models have been developed to describe Sn accounting for statistical cure: the asymptotic value of Sn is P [1]. Boussari et al. have developed a new cure model that allows a direct estimation of the time-to-cure by including the Time-to-Null-Excess-Hazard (TNEH) as a covariate-dependent parameter to be estimated [2]. In this model, P is the value of Sn when time elapsed since diagnosis equals TNEH. Objective: To compare the performances of the TNEH cure model, with that of a non-mixture flexible cure model and of a mixture cure model through a simulation study and applications to real datasets. Methods: Time-of-death was generated as the minimum of time-of-death due to other cause and due to cancer. This latter was consecutively simulated using mixture and TNEH cure models. Three different scenarios were considered according to the evolution of EH with time since diagnosis. Each scenario mimicked Sn from real situations: poor, medium, and good prognosis cancers. For each of the six situations, we generated 1,000 samples. We estimated P and Sn 3, 5 and 10 years after diagnosis through the three models with age-group as a covariate. Indicators of performances (bias, root-mean-square-error and coverage-rate) were studied by age group. Sn was estimated on three sets of corresponding real data (pancreatic, colon and testicular cancers) which originated from the French cancer registries database (FRANCIM). Results: The performances of the three models were correct in the situations in which EH reached zero. Unlike the other two, TNEH model performed poorly for the situations in which although EH became low, it did not reach zero. From real data, Sn curves from the three models were identical for testicular cancer. Sn estimated with TNEH differed from the others for colon and pancreatic cancers. Conclusions: The TNEH cure model can be used to estimate Sn if EH reaches zero. On this condition the model allows estimating the time-to-cure.