Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD): The TRIPOD Statement

Editors' Note: In order to encourage dissemination of the TRIPOD Statement, this article is freely accessible on the Annals of Internal Medicine Web site ( www.annals.org ) and will be also published in BJOG, British Journal of Cancer, British Journal of Surgery, BMC Medicine, British Medical Journal, Circulation, Diabetic Medicine, European Journal of Clinical Investigation, European Urology, and Journal of Clinical Epidemiology. The authors jointly hold the copyright of this article. An accompanying explanation and elaboration article is freely available only at www.annals.org ; Annals of Internal Medicine holds copyright for that article. In medicine, patients with their care providers are confronted with making numerous decisions on the basis of an estimated risk or probability that a specific disease or condition is present (diagnostic setting) or a specific event will occur in the future (prognostic setting) (Figure 1). In the diagnostic setting, the probability that a particular disease is present can be used, for example, to inform the referral of patients for further testing, initiate treatment directly, or reassure patients that a serious cause for their symptoms is unlikely. In the prognostic setting, predictions can be used for planning lifestyle or therapeutic decisions based on the risk for developing a particular outcome or state of health within a specific period (1, 2). Such estimates of risk can also be used to risk-stratify participants in therapeutic clinical trials (3, 4). Figure 1. Schematic representation of diagnostic and prognostic prediction modeling studies. The nature of the prediction in diagnosis is estimating the probability that a specific outcome or disease is present (or absent) within an individual, at this point in timethat is, the moment of prediction (T= 0). In prognosis, the prediction is about whether an individual will experience a specific event or outcome within a certain time period. In other words, in diagnostic prediction the interest is in principle a cross-sectional relationship, whereas prognostic prediction involves a longitudinal relationship. Nevertheless, in diagnostic modeling studies, for logistical reasons, a time window between predictor (index test) measurement and the reference standard is often necessary. Ideally, this interval should be as short as possible and without starting any treatment within this period. In both the diagnostic and prognostic setting, estimates of probabilities are rarely based on a single predictor (5). Doctors naturally integrate several patient characteristics and symptoms (predictors, test results) to make a prediction (see Figure 2 for differences in common terminology between diagnostic and prognostic studies). Prediction is therefore inherently multivariable. Prediction models (also commonly called prognostic models, risk scores, or prediction rules [6]) are tools that combine multiple predictors by assigning relative weights to each predictor to obtain a risk or probability (1, 2). Well-known prediction models include the Framingham Risk Score (7), Ottawa Ankle Rules (8), EuroScore (9), Nottingham Prognostic Index (10), and the Simplified Acute Physiology Score (11). Figure 2. Similarities and differences between diagnostic and prognostic prediction models. Prediction Model Studies Prediction model studies can be broadly categorized as model development (12), model validation (with or without updating) (13) or a combination of both (Figure 3). Model development studies aim to derive a prediction model by selecting the relevant predictors and combining them statistically into a multivariable model. Logistic and Cox regression are most frequently used for short-term (for example, disease absent vs. present, 30-day mortality) and long-term (for example, 10-year risk) outcomes, respectively (1214). Studies may also focus on quantifying the incremental or added predictive value of a specific predictor (for example, newly discovered) to a prediction model (18). Figure 3. Types of prediction model studies covered by the TRIPOD Statement. D = development data; V = validation data. Quantifying the predictive ability of a model on the same data from which the model was developed (often referred to as apparent performance) will tend to give an optimistic estimate of performance, owing to overfitting (too few outcome events relative to the number of candidate predictors) and the use of predictor selection strategies (19). Studies developing new prediction models should therefore always include some form of internal validation to quantify any optimism in the predictive performance (for example, calibration and discrimination) of the developed model. Internal validation techniques use only the original study sample and include such methods as bootstrapping or cross-validation. Internal validation is a necessary part of model development (2). Overfitting, optimism, and miscalibration may also be addressed and accounted for during the model development by applying shrinkage (for example, heuristic or based on bootstrapping techniques) or penalization procedures (for example, ridge regression or lasso) (20). After developing a prediction model, it is strongly recommended to evaluate the performance of the model in other participant data than was used for the model development. Such external validation requires that for each individual in the new data set, outcome predictions are made using the original model (that is, the published regression formula) and compared with the observed outcomes (13, 14). External validation may use participant data collected by the same investigators, typically using the same predictor and outcome definitions and measurements, but sampled from a later period (temporal or narrow validation); by other investigators in another hospital or country, sometimes using different definitions and measurements (geographic or broad validation); in similar participants but from an intentionally different setting (for example, model developed in secondary care and assessed in similar participants but selected from primary care); or even in other types of participants (for example, model developed in adults and assessed in children, or developed for predicting fatal events and assessed for predicting nonfatal events) (13, 15, 17, 21, 22). In case of poor performance, the model can be updated or adjusted on the basis of the validation data set (13). Reporting of Multivariable Prediction Model Studies Studies developing or validating a multivariable prediction model share specific challenges for researchers (6). Several reviews have evaluated the quality of published reports that describe the development or validation prediction models (2328). For example, Mallett and colleagues (26) examined 47 reports published in 2005 presenting new prediction models in cancer. Reporting was found to be poor, with insufficient information described in all aspects of model development, from descriptions of patient data to statistical modeling methods. Collins and colleagues (24) evaluated the methodological conduct and reporting of 39 reports published before May 2011 describing the development of models to predict prevalent or incident type 2 diabetes. Reporting was also found to be generally poor, with key details on which predictors were examined, the handling and reporting of missing data, and model-building strategy often poorly described. Bouwmeester and colleagues (23) evaluated 71 reports, published in 2008 in 6 high-impact general medical journals, and likewise observed an overwhelmingly poor level of reporting. These and other reviews provide a clear picture that, across different disease areas and different journals, there is a generally poor level of reporting of prediction model studies (6, 2327, 29). Furthermore, these reviews have shown that serious deficiencies in the statistical methods, use of small data sets, inappropriate handling of missing data, and lack of validation are common (6, 2327, 29). Such deficiencies ultimately lead to prediction models that are not or should not be used. It is therefore not surprising, and fortunate, that very few prediction models, relative to the large number of models published, are widely implemented or used in clinical practice (6). Prediction models in medicine have proliferated in recent years. Health care providers and policy makers are increasingly recommending the use of prediction models within clinical practice guidelines to inform decision making at various stages in the clinical pathway (30, 31). It is a general requirement of reporting of research that other researchers can, if required, replicate all the steps taken and obtain the same results (32). It is therefore essential that key details of how a prediction model was developed and validated be clearly reported to enable synthesis and critical appraisal of all relevant information (14, 3336). Reporting Guidelines for Prediction Model Studies: The TRIPOD Statement We describe the development of the TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis) Statement, a guideline specifically designed for the reporting of studies developing or validating a multivariable prediction model, whether for diagnostic or prognostic purposes. TRIPOD is not intended for multivariable modeling in etiologic studies or for studies investigating single prognostic factors (37). Furthermore, TRIPOD is also not intended for impact studies that quantify the impact of using a prediction model on participant or doctors' behavior and management, participant health outcomes, or cost-effectiveness of care, compared with not using the model (13, 38). Reporting guidelines for observational (the STrengthening the Reporting of OBservational studies in Epidemiology [STROBE]) (39), tumor marker (REporting recommendations for tumour MARKer prognostic studies [REMARK]) (37), diagnostic accuracy (STAndards f