Search

A myriad of measures to illustrate performance of predictive artificial intelligence (AI) models have been proposed in the literature. Selecting appropriate performance measures is essential for predictive AI models that are developed to be used in medical practice, because poorly performing models may harm patients and lead to increased costs. We aim to assess the merits of classic and contemporary performance measures when validating predictive AI models for use in medical practice. We focus on models with a binary outcome. We discuss 32 performance measures covering five performance domains (discrimination, calibration, overall, classification, and clinical utility) along with accompanying graphical assessments. The first four domains cover statistical performance, the fifth domain covers decision-analytic performance. We explain why two key characteristics are important when selecting which performance measures to assess: (1) whether the measure's expected value is optimized when it is calculated using the correct probabilities (i.e., a "proper" measure), and (2) whether they reflect either purely statistical performance or decision-analytic performance by properly considering misclassification costs. Seventeen measures exhibit both characteristics, fourteen measures exhibited one characteristic, and one measure possessed neither characteristic (the F1 measure). All classification measures (such as classification accuracy and F1) are improper for clinically relevant decision thresholds other than 0.5 or the prevalence. We recommend the following measures and plots as essential to report: AUROC, calibration plot, a clinical utility measure such as net benefit with decision curve analysis, and a plot with probability distributions per outcome category., Comment: 60 pages, 8 tables, 11 figures, two supplementary appendices

When evaluating the performance of a model for individualised risk prediction, the sample size needs to be large enough to precisely estimate the performance measures of interest. Current sample size guidance is based on precisely estimating calibration, discrimination, and net benefit, which should be the first stage of calculating the minimum required sample size. However, when a clinically important threshold is used for classification, other performance measures can also be used. We extend the previously published guidance to precisely estimate threshold-based performance measures. We have developed closed-form solutions to estimate the sample size required to target sufficiently precise estimates of accuracy, specificity, sensitivity, PPV, NPV, and F1-score in an external evaluation study of a prediction model with a binary outcome. This approach requires the user to pre-specify the target standard error and the expected value for each performance measure. We describe how the sample size formulae were derived and demonstrate their use in an example. Extension to time-to-event outcomes is also considered. In our examples, the minimum sample size required was lower than that required to precisely estimate the calibration slope, and we expect this would most often be the case. Our formulae, along with corresponding Python code and updated R and Stata commands (pmvalsampsize), enable researchers to calculate the minimum sample size needed to precisely estimate threshold-based performance measures in an external evaluation study. These criteria should be used alongside previously published criteria to precisely estimate the calibration, discrimination, and net-benefit., Comment: 27 pages, 1 figure

Risk prediction models are increasingly used in healthcare to aid in clinical decision making. In most clinical contexts, model calibration (i.e., assessing the reliability of risk estimates) is critical. Data available for model development are often not perfectly balanced with respect to the modeled outcome (i.e., individuals with vs. without the event of interest are not equally represented in the data). It is common for researchers to correct this class imbalance, yet, the effect of such imbalance corrections on the calibration of machine learning models is largely unknown. We studied the effect of imbalance corrections on model calibration for a variety of machine learning algorithms. Using extensive Monte Carlo simulations we compared the out-of-sample predictive performance of models developed with an imbalance correction to those developed without a correction for class imbalance across different data-generating scenarios (varying sample size, the number of predictors and event fraction). Our findings were illustrated in a case study using MIMIC-III data. In all simulation scenarios, prediction models developed without a correction for class imbalance consistently had equal or better calibration performance than prediction models developed with a correction for class imbalance. The miscalibration introduced by correcting for class imbalance was characterized by an over-estimation of risk and was not always able to be corrected with re-calibration. Correcting for class imbalance is not always necessary and may even be harmful for clinical prediction models which aim to produce reliable risk estimates on an individual basis.

Validation metrics are key for the reliable tracking of scientific progress and for bridging the current chasm between artificial intelligence (AI) research and its translation into practice. However, increasing evidence shows that particularly in image analysis, metrics are often chosen inadequately in relation to the underlying research problem. This could be attributed to a lack of accessibility of metric-related knowledge: While taking into account the individual strengths, weaknesses, and limitations of validation metrics is a critical prerequisite to making educated choices, the relevant knowledge is currently scattered and poorly accessible to individual researchers. Based on a multi-stage Delphi process conducted by a multidisciplinary expert consortium as well as extensive community feedback, the present work provides the first reliable and comprehensive common point of access to information on pitfalls related to validation metrics in image analysis. Focusing on biomedical image analysis but with the potential of transfer to other fields, the addressed pitfalls generalize across application domains and are categorized according to a newly created, domain-agnostic taxonomy. To facilitate comprehension, illustrations and specific examples accompany each pitfall. As a structured body of information accessible to researchers of all levels of expertise, this work enhances global comprehension of a key topic in image analysis validation., Comment: Shared first authors: Annika Reinke and Minu D. Tizabi; shared senior authors: Lena Maier-Hein and Paul F. J\"ager. Published in Nature Methods. arXiv admin note: text overlap with arXiv:2206.01653

Objective: Text mining of clinical notes embedded in electronic medical records is increasingly used to extract patient characteristics otherwise not or only partly available, to assess their association with relevant health outcomes. As manual data labeling needed to develop text mining models is resource intensive, we investigated whether off-the-shelf text mining models developed at external institutions, together with limited within-institution labeled data, could be used to reliably extract study variables to conduct association studies. Materials and Methods: We developed multiple text mining models on different combinations of within-institution and external-institution data to extract social factors from discharge reports of intensive care patients. Subsequently, we assessed the associations between social factors and having a do-not-resuscitate/intubate code. Results: Important differences were found between associations based on manually labeled data compared to text-mined social factors in three out of five cases. Adopting external-institution text mining models using manually labeled within-institution data resulted in models with higher F1-scores, but not in meaningfully different associations. Discussion: While text mining facilitated scaling analyses to larger samples leading to discovering a larger number of associations, the estimates may be unreliable. Confirmation is needed with better text mining models, ideally on a larger manually labeled dataset. Conclusion: The currently used text mining models were not sufficiently accurate to be used reliably in an association study. Model adaptation using within-institution data did not improve the estimates. Further research is needed to set conditions for reliable use of text mining in medical research.

Multinomial logistic regression models allow one to predict the risk of a categorical outcome with more than 2 categories. When developing such a model, researchers should ensure the number of participants (n) is appropriate relative to the number of events (E.k) and the number of predictor parameters (p.k) for each category k. We propose three criteria to determine the minimum n required in light of existing criteria developed for binary outcomes. The first criteria aims to minimise the model overfitting. The second aims to minimise the difference between the observed and adjusted R2 Nagelkerke. The third criterion aims to ensure the overall risk is estimated precisely. For criterion (i), we show the sample size must be based on the anticipated Cox-snell R2 of distinct one-to-one logistic regression models corresponding to the sub-models of the multinomial logistic regression, rather than on the overall Cox-snell R2 of the multinomial logistic regression. We tested the performance of the proposed criteria (i) through a simulation study, and found that it resulted in the desired level of overfitting. Criterion (ii) and (iii) are natural extensions from previously proposed criteria for binary outcomes. We illustrate how to implement the sample size criteria through a worked example considering the development of a multinomial risk prediction model for tumour type when presented with an ovarian mass. Code is provided for the simulation and worked example. We will embed our proposed criteria within the pmsampsize R library and Stata modules.

Background: Existing guidelines for handling missing data are generally not consistent with the goals of prediction modelling, where missing data can occur at any stage of the model pipeline. Multiple imputation (MI), often heralded as the gold standard approach, can be challenging to apply in the clinic. Clearly, the outcome cannot be used to impute data at prediction time. Regression imputation (RI) may offer a pragmatic alternative in the prediction context, that is simpler to apply in the clinic. Moreover, the use of missing indicators can handle informative missingness, but it is currently unknown how well they perform within CPMs. Methods: We performed a simulation study where data were generated under various missing data mechanisms to compare the predictive performance of CPMs developed using both imputation methods. We consider deployment scenarios where missing data is permitted/prohibited, and develop models that use/omit the outcome during imputation and include/omit missing indicators. Results: When complete data must be available at deployment, our findings were in line with widely used recommendations; that the outcome should be used to impute development data under MI, yet omitted under RI. When imputation is applied at deployment, omitting the outcome from the imputation at development was preferred. Missing indicators improved model performance in some specific cases, but can be harmful when missingness is dependent on the outcome. Conclusion: We provide evidence that commonly taught principles of handling missing data via MI may not apply to CPMs, particularly when data can be missing at deployment. In such settings, RI and missing indicator methods can (marginally) outperform MI. As shown, the performance of the missing data handling method must be evaluated on a study-by-study basis, and should be based on whether missing data are allowed at deployment., Comment: 42 pages. Submitted to Statistical Methods in Medical Research in October 2021

Increasing evidence shows that flaws in machine learning (ML) algorithm validation are an underestimated global problem. Particularly in automatic biomedical image analysis, chosen performance metrics often do not reflect the domain interest, thus failing to adequately measure scientific progress and hindering translation of ML techniques into practice. To overcome this, our large international expert consortium created Metrics Reloaded, a comprehensive framework guiding researchers in the problem-aware selection of metrics. Following the convergence of ML methodology across application domains, Metrics Reloaded fosters the convergence of validation methodology. The framework was developed in a multi-stage Delphi process and is based on the novel concept of a problem fingerprint - a structured representation of the given problem that captures all aspects that are relevant for metric selection, from the domain interest to the properties of the target structure(s), data set and algorithm output. Based on the problem fingerprint, users are guided through the process of choosing and applying appropriate validation metrics while being made aware of potential pitfalls. Metrics Reloaded targets image analysis problems that can be interpreted as a classification task at image, object or pixel level, namely image-level classification, object detection, semantic segmentation, and instance segmentation tasks. To improve the user experience, we implemented the framework in the Metrics Reloaded online tool, which also provides a point of access to explore weaknesses, strengths and specific recommendations for the most common validation metrics. The broad applicability of our framework across domains is demonstrated by an instantiation for various biological and medical image analysis use cases., Comment: Shared first authors: Lena Maier-Hein, Annika Reinke. arXiv admin note: substantial text overlap with arXiv:2104.05642 Published in Nature Methods

Methods to correct class imbalance, i.e. imbalance between the frequency of outcome events and non-events, are receiving increasing interest for developing prediction models. We examined the effect of imbalance correction on the performance of standard and penalized (ridge) logistic regression models in terms of discrimination, calibration, and classification. We examined random undersampling, random oversampling and SMOTE using Monte Carlo simulations and a case study on ovarian cancer diagnosis. The results indicated that all imbalance correction methods led to poor calibration (strong overestimation of the probability to belong to the minority class), but not to better discrimination in terms of the area under the receiver operating characteristic curve. Imbalance correction improved classification in terms of sensitivity and specificity, but similar results were obtained by shifting the probability threshold instead. Our study shows that outcome imbalance is not a problem in itself, and that imbalance correction may even worsen model performance., Comment: Main paper 21 pages, Supplement 53 pages

Calibration is a vital aspect of the performance of risk prediction models, but research in the context of ordinal outcomes is scarce. This study compared calibration measures for risk models predicting a discrete ordinal outcome, and investigated the impact of the proportional odds assumption on calibration and overfitting. We studied the multinomial, cumulative, adjacent category, continuation ratio, and stereotype logit/logistic models. To assess calibration, we investigated calibration intercepts and slopes, calibration plots, and the estimated calibration index. Using large sample simulations, we studied the performance of models for risk estimation under various conditions, assuming that the true model has either a multinomial logistic form or a cumulative logit proportional odds form. Small sample simulations were used to compare the tendency for overfitting between models. As a case study, we developed models to diagnose the degree of coronary artery disease (five categories) in symptomatic patients. When the true model was multinomial logistic, proportional odds models often yielded poor risk estimates, with calibration slopes deviating considerably from unity even on large model development datasets. The stereotype logistic model improved the calibration slope, but still provided biased risk estimates for individual patients. When the true model had a cumulative logit proportional odds form, multinomial logistic regression provided biased risk estimates, although these biases were modest. Non-proportional odds models require more parameters to be estimated from the data, and hence suffered more from overfitting. Despite larger sample size requirements, we generally recommend multinomial logistic regression for risk prediction modeling of discrete ordinal outcomes., Comment: Revised version submitted to Statistics in Medicine

While the importance of automatic image analysis is continuously increasing, recent meta-research revealed major flaws with respect to algorithm validation. Performance metrics are particularly key for meaningful, objective, and transparent performance assessment and validation of the used automatic algorithms, but relatively little attention has been given to the practical pitfalls when using specific metrics for a given image analysis task. These are typically related to (1) the disregard of inherent metric properties, such as the behaviour in the presence of class imbalance or small target structures, (2) the disregard of inherent data set properties, such as the non-independence of the test cases, and (3) the disregard of the actual biomedical domain interest that the metrics should reflect. This living dynamically document has the purpose to illustrate important limitations of performance metrics commonly applied in the field of image analysis. In this context, it focuses on biomedical image analysis problems that can be phrased as image-level classification, semantic segmentation, instance segmentation, or object detection task. The current version is based on a Delphi process on metrics conducted by an international consortium of image analysis experts from more than 60 institutions worldwide., Comment: Shared first authors: Annika Reinke and Minu D. Tizabi. This is a dynamic paper on limitations of commonly used metrics. It discusses metrics for image-level classification, semantic and instance segmentation, and object detection. For missing use cases, comments or questions, please contact a.reinke@dkfz.de. Substantial contributions to this document will be acknowledged with a co-authorship

Measurement error in a covariate or the outcome of regression models is common, but is often ignored, even though measurement error can lead to substantial bias in the estimated covariate-outcome association. While several texts on measurement error correction methods are available, these methods remain seldomly applied. To improve the use of measurement error correction methodology, we developed mecor, an R package that implements measurement error correction methods for regression models with continuous outcomes. Measurement error correction requires information about the measurement error model and its parameters. This information can be obtained from four types of studies, used to estimate the parameters of the measurement error model: an internal validation study, a replicates study, a calibration study and an external validation study. In the package mecor, regression calibration methods and a maximum likelihood method are implemented to correct for measurement error in a continuous covariate in regression analyses. Additionally, methods of moments methods are implemented to correct for measurement error in the continuous outcome in regression analyses. Variance estimation of the corrected estimators is provided in closed form and using the bootstrap., Comment: 34 pages (including appendix), software package

Clinical prediction models are developed widely across medical disciplines. When predictors in such models are highly collinear, unexpected or spurious predictor-outcome associations may occur, thereby potentially reducing face-validity and explainability of the prediction model. Collinearity can be dealt with by exclusion of collinear predictors, but when there is no a priori motivation (besides collinearity) to include or exclude specific predictors, such an approach is arbitrary and possibly inappropriate. We compare different methods to address collinearity, including shrinkage, dimensionality reduction, and constrained optimization. The effectiveness of these methods is illustrated via simulations. In the conducted simulations, no effect of collinearity was observed on predictive outcomes. However, a negative effect of collinearity on the stability of predictor selection was found, affecting all compared methods, but in particular methods that perform strong predictor selection (e.g., Lasso).}

In observational research treatment effects, the average treatment effect (ATE) estimator may be biased if a confounding variable is misclassified. We discuss the impact of classification error in a dichotomous confounding variable in analyses using marginal structural models estimated using inverse probability weighting (MSMs-IPW) and compare this with its impact in conditional regression models, focusing on a point-treatment study with a continuous outcome. Expressions were derived for the bias in the ATE estimator from a MSM-IPW and conditional model by using the potential outcome framework. Based on these expressions, we propose a sensitivity analysis to investigate and quantify the bias due to classification error in a confounding variable in MSMs-IPW. Compared to bias in the ATE estimator from a conditional model, the bias in MSM-IPW can be dissimilar in magnitude but the bias will always be equal in sign. A simulation study was conducted to study the finite sample performance of MSMs-IPW and conditional models if a confounding variable is misclassified. Simulation results showed that confidence intervals of the treatment effect obtained from MSM-IPW are generally wider and coverage of the true treatment effect is higher compared to a conditional model, ranging from over coverage if there is no classification error to smaller under coverage when there is classification error. The use of the bias expressions to inform a sensitivity analysis was demonstrated in a study of blood pressure lowering therapy. It is important to consider the potential impact of classification error in a confounding variable in studies of treatment effects and a sensitivity analysis provides an opportunity to quantify the impact of such errors on causal conclusions. An online tool for sensitivity analyses was developed: https://lindanab.shinyapps.io/SensitivityAnalysis., Comment: 25 pages, 3 figures, 3 tables

When developing risk prediction models, shrinkage methods are recommended, especially when the sample size is limited. Several earlier studies have shown that the shrinkage of model coefficients can reduce overfitting of the prediction model and subsequently result in better predictive performance on average. In this simulation study, we aimed to investigate the variability of regression shrinkage on predictive performance for a binary outcome, with focus on the calibration slope. The slope indicates whether risk predictions are too extreme (slope < 1) or not extreme enough (slope > 1). We investigated the following shrinkage methods in comparison to standard maximum likelihood estimation: uniform shrinkage (likelihood-based and bootstrap-based), ridge regression, penalized maximum likelihood, LASSO regression, adaptive LASSO, non-negative garrote, and Firth's correction. There were three main findings. First, shrinkage improved calibration slopes on average. Second, the between-sample variability of calibration slopes was often increased relative to maximum likelihood. Among the shrinkage methods, the bootstrap-based uniform shrinkage worked well overall. In contrast to other shrinkage approaches, Firth's correction had only a small shrinkage effect but did so with low variability. Third, the correlation between the estimated shrinkage and the optimal shrinkage to remove overfitting was typically negative. Hence, although shrinkage improved predictions on average, it often worked poorly in individual datasets, in particular when shrinkage was most needed. The observed variability of shrinkage methods implies that these methods do not solve problems associated with small sample size or low number of events per variable., Comment: 138 pages (incl 114 supplementary pages). Main document: 5 figures and 2 tables

Joint misclassification of exposure and outcome variables can lead to considerable bias in epidemiological studies of causal exposure-outcome effects. In this paper, we present a new maximum likelihood based estimator for the marginal causal odd-ratio that simultaneously adjusts for confounding and several forms of joint misclassification of the exposure and outcome variables. The proposed method relies on validation data for the construction of weights that account for both sources of bias. The weighting estimator, which is an extension of the exposure misclassification weighting estimator proposed by Gravel and Platt (Statistics in Medicine, 2018), is applied to reinfarction data. Simulation studies were carried out to study its finite sample properties and compare it with methods that do not account for confounding or misclassification. The new estimator showed favourable large sample properties in the simulations. Further research is needed to study the sensitivity of the proposed method and that of alternatives to violations of their assumptions. The implementation of the estimator is facilitated by a new R function in an existing R package., Comment: 36 pages, 7 tables, 1 figure

In randomised trials, continuous endpoints are often measured with some degree of error. This study explores the impact of ignoring measurement error, and proposes methods to improve statistical inference in the presence of measurement error. Three main types of measurement error in continuous endpoints are considered: classical, systematic and differential. For each measurement error type, a corrected effect estimator is proposed. The corrected estimators and several methods for confidence interval estimation are tested in a simulation study. These methods combine information about error-prone and error-free measurements of the endpoint in individuals not included in the trial (external calibration sample). We show that if measurement error in continuous endpoints is ignored, the treatment effect estimator is unbiased when measurement error is classical, while Type-II error is increased at a given sample size. Conversely, the estimator can be substantially biased when measurement error is systematic or differential. In those cases, bias can largely be prevented and inferences improved upon using information from an external calibration sample, of which the required sample size increases as the strength of the association between the error-prone and error-free endpoint decreases. Measurement error correction using already a small (external) calibration sample is shown to improve inferences and should be considered in trials with error-prone endpoints. Implementation of the proposed correction methods is accommodated by a new software package for R., Comment: 37 pages, 4 figures, 3 tables

Data mining and machine learning techniques such as classification and regression trees (CART) represent a promising alternative to conventional logistic regression for propensity score estimation. Whereas incomplete data preclude the fitting of a logistic regression on all subjects, CART is appealing in part because some implementations allow for incomplete records to be incorporated in the tree fitting and provide propensity score estimates for all subjects. Based on theoretical considerations, we argue that the automatic handling of missing data by CART may however not be appropriate. Using a series of simulation experiments, we examined the performance of different approaches to handling missing covariate data; (i) applying the CART algorithm directly to the (partially) incomplete data, (ii) complete case analysis, and (iii) multiple imputation. Performance was assessed in terms of bias in estimating exposure-outcome effects \add{among the exposed}, standard error, mean squared error and coverage. Applying the CART algorithm directly to incomplete data resulted in bias, even in scenarios where data were missing completely at random. Overall, multiple imputation followed by CART resulted in the best performance. Our study showed that automatic handling of missing data in CART can cause serious bias and does not outperform multiple imputation as a means to account for missing data., Comment: 29 pages, 5 tables

It is widely acknowledged that the predictive performance of clinical prediction models should be studied in patients that were not part of the data in which the model was derived. Out-of-sample performance can be hampered when predictors are measured differently at derivation and external validation. This may occur, for instance, when predictors are measured using different measurement protocols or when tests are produced by different manufacturers. Although such heterogeneity in predictor measurement between deriviation and validation data is common, the impact on the out-of-sample performance is not well studied. Using analytical and simulation approaches, we examined out-of-sample performance of prediction models under various scenarios of heterogeneous predictor measurement. These scenarios were defined and clarified using an established taxonomy of measurement error models. The results of our simulations indicate that predictor measurement heterogeneity can induce miscalibration of prediction and affects discrimination and overall predictive accuracy, to extents that the prediction model may no longer be considered clinically useful. The measurement error taxonomy was found to be helpful in identifying and predicting effects of heterogeneous predictor measurements between settings of prediction model derivation and validation. Our work indicates that homogeneity of measurement strategies across settings is of paramount importance in prediction research., Comment: 32 pages, 4 figures