Dataset Information

External validation of multivariable prediction models: a systematic review of methodological conduct and reporting.

ABSTRACT:

Background

Before considering whether to use a multivariable (diagnostic or prognostic) prediction model, it is essential that its performance be evaluated in data that were not used to develop the model (referred to as external validation). We critically appraised the methodological conduct and reporting of external validation studies of multivariable prediction models.

Methods

We conducted a systematic review of articles describing some form of external validation of one or more multivariable prediction models indexed in PubMed core clinical journals published in 2010. Study data were extracted in duplicate on design, sample size, handling of missing data, reference to the original study developing the prediction models and predictive performance measures.

Results

11,826 articles were identified and 78 were included for full review, which described the evaluation of 120 prediction models. in participant data that were not used to develop the model. Thirty-three articles described both the development of a prediction model and an evaluation of its performance on a separate dataset, and 45 articles described only the evaluation of an existing published prediction model on another dataset. Fifty-seven percent of the prediction models were presented and evaluated as simplified scoring systems. Sixteen percent of articles failed to report the number of outcome events in the validation datasets. Fifty-four percent of studies made no explicit mention of missing data. Sixty-seven percent did not report evaluating model calibration whilst most studies evaluated model discrimination. It was often unclear whether the reported performance measures were for the full regression model or for the simplified models.

Conclusions

The vast majority of studies describing some form of external validation of a multivariable prediction model were poorly reported with key details frequently not presented. The validation studies were characterised by poor design, inappropriate handling and acknowledgement of missing data and one of the most key performance measures of prediction models i.e. calibration often omitted from the publication. It may therefore not be surprising that an overwhelming majority of developed prediction models are not used in practice, when there is a dearth of well-conducted and clearly reported (external validation) studies describing their performance on independent participant data.

SUBMITTER: Collins GS

PROVIDER: S-EPMC3999945 | biostudies-literature |

REPOSITORIES: biostudies-literature

ACCESS DATA

Similar Datasets

Project description:Background and purposePrediction models for outcome of patients with acute ischemic stroke who will undergo endovascular treatment have been developed to improve patient management. The aim of the current study is to provide an overview of preintervention models for functional outcome after endovascular treatment and to validate these models with data from daily clinical practice.MethodsWe systematically searched within Medline, Embase, Cochrane, Web of Science, to include prediction models. Models identified from the search were validated in the MR CLEAN (Multicenter Randomized Clinical Trial of Endovascular Treatment for Acute Ischemic Stroke in the Netherlands) registry, which includes all patients treated with endovascular treatment within 6.5 hours after stroke onset in the Netherlands between March 2014 and November 2017. Predictive performance was evaluated according to discrimination (area under the curve) and calibration (slope and intercept of the calibration curve). Good functional outcome was defined as a score of 0-2 or 0-3 on the modified Rankin Scale depending on the model.ResultsAfter screening 3468 publications, 19 models were included in this validation. Variables included in the models mainly addressed clinical and imaging characteristics at baseline. In the validation cohort of 3156 patients, discriminative performance ranged from 0.61 (SPAN-100 [Stroke Prognostication Using Age and NIH Stroke Scale]) to 0.80 (MR PREDICTS). Best-calibrated models were THRIVE (The Totaled Health Risks in Vascular Events; intercept -0.06 [95% CI, -0.14 to 0.02]; slope 0.84 [95% CI, 0.75-0.95]), THRIVE-c (intercept 0.08 [95% CI, -0.02 to 0.17]; slope 0.71 [95% CI, 0.65-0.77]), Stroke Checkerboard score (intercept -0.05 [95% CI, -0.13 to 0.03]; slope 0.97 [95% CI, 0.88-1.08]), and MR PREDICTS (intercept 0.43 [95% CI, 0.33-0.52]; slope 0.93 [95% CI, 0.85-1.01]).ConclusionsThe THRIVE-c score and MR PREDICTS both showed a good combination of discrimination and calibration and were, therefore, superior in predicting functional outcome for patients with ischemic stroke after endovascular treatment within 6.5 hours. Since models used different predictors and several models had relatively good predictive performance, the decision on which model to use in practice may also depend on simplicity of the model, data availability, and the comparability of the population and setting.

Project description:ObjectiveTo validate all diagnostic prediction models for ruling out pulmonary embolism that are easily applicable in primary care.DesignSystematic review followed by independent external validation study to assess transportability of retrieved models to primary care medicine.Setting300 general practices in the Netherlands.ParticipantsIndividual patient dataset of 598 patients with suspected acute pulmonary embolism in primary care.Main outcome measuresDiscriminative ability of all models retrieved by systematic literature search, assessed by calculation and comparison of C statistics. After stratification into groups with high and low probability of pulmonary embolism according to pre-specified model cut-offs combined with qualitative D-dimer test, sensitivity, specificity, efficiency (overall proportion of patients with low probability of pulmonary embolism), and failure rate (proportion of pulmonary embolism cases in group of patients with low probability) were calculated for all models.ResultsTen published prediction models for the diagnosis of pulmonary embolism were found. Five of these models could be validated in the primary care dataset: the original Wells, modified Wells, simplified Wells, revised Geneva, and simplified revised Geneva models. Discriminative ability was comparable for all models (range of C statistic 0.75-0.80). Sensitivity ranged from 88% (simplified revised Geneva) to 96% (simplified Wells) and specificity from 48% (revised Geneva) to 53% (simplified revised Geneva). Efficiency of all models was between 43% and 48%. Differences were observed between failure rates, especially between the simplified Wells and the simplified revised Geneva models (failure rates 1.2% (95% confidence interval 0.2% to 3.3%) and 3.1% (1.4% to 5.9%), respectively; absolute difference -1.98% (-3.33% to -0.74%)). Irrespective of the diagnostic prediction model used, three patients were incorrectly classified as having low probability of pulmonary embolism; pulmonary embolism was diagnosed only after referral to secondary care.ConclusionsFive diagnostic pulmonary embolism prediction models that are easily applicable in primary care were validated in this setting. Whereas efficiency was comparable for all rules, the Wells rules gave the best performance in terms of lower failure rates.

Project description:ObjectiveTo systematically review the conduct and reporting of formula trials.DesignSystematic review.Data sourcesMedline, Embase, and Cochrane Central Register of Controlled Trials (CENTRAL) were searched from 1 January 2006 to 31 December 2020.Review methodsIntervention trials comparing at least two formula products in children less than three years of age were included, but not trials of human breast milk or fortifiers of breast milk. Data were extracted in duplicate and primary outcome data were synthesised for meta-analysis with a random effects model weighted by the inverse variance method. Risk of bias was evaluated with Cochrane risk of bias version 2.0, and risk of undermining breastfeeding was evaluated according to published consensus guidance. Primary outcomes of the trials included in the systematic review were identified from clinical trial registries, protocols, or trial publications.Results22 201 titles were screened and 307 trials were identified that were published between 2006 and 2020, of which 73 (24%) trials in 13 197 children were prospectively registered. Another 111 unpublished but registered trials in 17 411 children were identified. Detailed analysis was undertaken for 125 trials (23 757 children) published since 2015. Seventeen (14%) of these recently published trials were conducted independently of formula companies, 26 (21%) were prospectively registered with a clear aim and primary outcome, and authors or sponsors shared prospective protocols for 11 (9%) trials. Risk of bias was low in five (4%) and high in 100 (80%) recently published trials, mainly because of inappropriate exclusions from analysis and selective reporting. For 68 recently published superiority trials, a pooled standardised mean difference of 0.51 (range -0.43 to 3.29) was calculated with an asymmetrical funnel plot (Egger's test P<0.001), which reduced to 0.19 after correction for asymmetry. Primary outcomes were reported by authors as favourable in 86 (69%) trials, and 115 (92%) abstract conclusions were favourable. One of 38 (3%) trials in partially breastfed infants reported adequate support for breastfeeding and 14 of 87 (16%) trials in non-breastfed infants confirmed the decision not to breastfeed was firmly established before enrolment in the trial.ConclusionsThe results show that formula trials lack independence or transparency, and published outcomes are biased by selective reporting.Systematic review registrationPROSPERO 2018 CRD42018091928.

Project description:ObjectivesTo identify and assess the quality and accuracy of prognostic models for nephropathy and to validate these models in external cohorts of people with type 2 diabetes.DesignSystematic review and external validation.Data sourcesPubMed and Embase.Eligibility criteriaStudies describing the development of a model to predict the risk of nephropathy, applicable to people with type 2 diabetes.MethodsScreening, data extraction, and risk of bias assessment were done in duplicate. Eligible models were externally validated in the Hoorn Diabetes Care System (DCS) cohort (n=11 450) for the same outcomes for which they were developed. Risks of nephropathy were calculated and compared with observed risk over 2, 5, and 10 years of follow-up. Model performance was assessed based on intercept adjusted calibration and discrimination (Harrell's C statistic).Results41 studies included in the systematic review reported 64 models, 46 of which were developed in a population with diabetes and 18 in the general population including diabetes as a predictor. The predicted outcomes included albuminuria, diabetic kidney disease, chronic kidney disease (general population), and end stage renal disease. The reported apparent discrimination of the 46 models varied considerably across the different predicted outcomes, from 0.60 (95% confidence interval 0.56 to 0.64) to 0.99 (not available) for the models developed in a diabetes population and from 0.59 (not available) to 0.96 (0.95 to 0.97) for the models developed in the general population. Calibration was reported in 31 of the 41 studies, and the models were generally well calibrated. 21 of the 64 retrieved models were externally validated in the Hoorn DCS cohort for predicting risk of albuminuria, diabetic kidney disease, and chronic kidney disease, with considerable variation in performance across prediction horizons and models. For all three outcomes, however, at least two models had C statistics >0.8, indicating excellent discrimination. In a secondary external validation in GoDARTS (Genetics of Diabetes Audit and Research in Tayside Scotland), models developed for diabetic kidney disease outperformed those for chronic kidney disease. Models were generally well calibrated across all three prediction horizons.ConclusionsThis study identified multiple prediction models to predict albuminuria, diabetic kidney disease, chronic kidney disease, and end stage renal disease. In the external validation, discrimination and calibration for albuminuria, diabetic kidney disease, and chronic kidney disease varied considerably across prediction horizons and models. For each outcome, however, specific models showed good discrimination and calibration across the three prediction horizons, with clinically accessible predictors, making them applicable in a clinical setting.Systematic review registrationPROSPERO CRD42020192831.

Project description:IntroductionPrediction models may contribute to personalized risk-based management of women at high risk of spontaneous preterm delivery. Although prediction models are published frequently, often with promising results, external validation generally is lacking. We performed a systematic review of prediction models for the risk of spontaneous preterm birth based on routine clinical parameters. Additionally, we externally validated and evaluated the clinical potential of the models.Material and methodsPrediction models based on routinely collected maternal parameters obtainable during first 16 weeks of gestation were eligible for selection. Risk of bias was assessed according to the CHARMS guidelines. We validated the selected models in a Dutch multicenter prospective cohort study comprising 2614 unselected pregnant women. Information on predictors was obtained by a web-based questionnaire. Predictive performance of the models was quantified by the area under the receiver operating characteristic curve (AUC) and calibration plots for the outcomes spontaneous preterm birth <37 weeks and <34 weeks of gestation. Clinical value was evaluated by means of decision curve analysis and calculating classification accuracy for different risk thresholds.ResultsFour studies describing five prediction models fulfilled the eligibility criteria. Risk of bias assessment revealed a moderate to high risk of bias in three studies. The AUC of the models ranged from 0.54 to 0.67 and from 0.56 to 0.70 for the outcomes spontaneous preterm birth <37 weeks and <34 weeks of gestation, respectively. A subanalysis showed that the models discriminated poorly (AUC 0.51-0.56) for nulliparous women. Although we recalibrated the models, two models retained evidence of overfitting. The decision curve analysis showed low clinical benefit for the best performing models.ConclusionsThis review revealed several reporting and methodological shortcomings of published prediction models for spontaneous preterm birth. Our external validation study indicated that none of the models had the ability to predict spontaneous preterm birth adequately in our population. Further improvement of prediction models, using recent knowledge about both model development and potential risk factors, is necessary to provide an added value in personalized risk assessment of spontaneous preterm birth.

Project description:BackgroundModeling count and binary data collected in hierarchical designs have increased the use of Generalized Linear Mixed Models (GLMMs) in medicine. This article presents a systematic review of the application and quality of results and information reported from GLMMs in the field of clinical medicine.MethodsA search using the Web of Science database was performed for published original articles in medical journals from 2000 to 2012. The search strategy included the topic "generalized linear mixed models","hierarchical generalized linear models", "multilevel generalized linear model" and as a research domain we refined by science technology. Papers reporting methodological considerations without application, and those that were not involved in clinical medicine or written in English were excluded.ResultsA total of 443 articles were detected, with an increase over time in the number of articles. In total, 108 articles fit the inclusion criteria. Of these, 54.6% were declared to be longitudinal studies, whereas 58.3% and 26.9% were defined as repeated measurements and multilevel design, respectively. Twenty-two articles belonged to environmental and occupational public health, 10 articles to clinical neurology, 8 to oncology, and 7 to infectious diseases and pediatrics. The distribution of the response variable was reported in 88% of the articles, predominantly Binomial (n = 64) or Poisson (n = 22). Most of the useful information about GLMMs was not reported in most cases. Variance estimates of random effects were described in only 8 articles (9.2%). The model validation, the method of covariate selection and the method of goodness of fit were only reported in 8.0%, 36.8% and 14.9% of the articles, respectively.ConclusionsDuring recent years, the use of GLMMs in medical literature has increased to take into account the correlation of data when modeling qualitative data or counts. According to the current recommendations, the quality of reporting has room for improvement regarding the characteristics of the analysis, estimation method, validation, and selection of the model.