Untapped potential of multicenter studies: a review of cardiovascular risk prediction models revealed inappropriate analyses and wide variation in reporting

The Impact of Multi-Institution Datasets on the Generalizability of Machine Learning Prediction Models in the ICU

OBJECTIVES

To evaluate the transferability of deep learning (DL) models for the early detection of adverse events to previously unseen hospitals.

DESIGN

Retrospective observational cohort study utilizing harmonized intensive care data from four public datasets.

SETTING

ICUs across Europe and the United States.

PATIENTS

Adult patients admitted to the ICU for at least 6 hours who had good data quality.

INTERVENTIONS

None.

MEASUREMENTS AND MAIN RESULTS

Using carefully harmonized data from a total of 334,812 ICU stays, we systematically assessed the transferability of DL models for three common adverse events: death, acute kidney injury (AKI), and sepsis. We tested whether using more than one data source and/or algorithmically optimizing for generalizability during training improves model performance at new hospitals. We found that models achieved high area under the receiver operating characteristic (AUROC) for mortality (0.838-0.869), AKI (0.823-0.866), and sepsis (0.749-0.824) at the training hospital. As expected, AUROC dropped when models were applied at other hospitals, sometimes by as much as -0.200. Using more than one dataset for training mitigated the performance drop, with multicenter models performing roughly on par with the best single-center model. Dedicated methods promoting generalizability did not noticeably improve performance in our experiments.

CONCLUSIONS

Our results emphasize the importance of diverse training data for DL-based risk prediction. They suggest that as data from more hospitals become available for training, models may become increasingly generalizable. Even so, good performance at a new hospital still depended on the inclusion of compatible hospitals during training.

Tips and tricks for successfully conducting a multicenter study

Multicenter studies play a crucial role in medical research and advancement, facilitating the application of new knowledge to clinical practice. These studies are associated with multiple benefits but are more complex than those involving a single center. With the philosophy that most of the qualities required to lead a multicenter study depend on attitude and can be learned, developed, and improved, in this manuscript, we share with the reader a series of recommendations that we consider important for successfully conducting such studies. The tips and tricks that will be discussed in detail are as follows: effectively leading the project; clearly defining viable and relevant objectives; designing a clear and detailed protocol; carefully selecting centers and collaborating investigators; meticulously designing the case report form; centrally managing the project efficiently; maintaining fluent communication with investigators; and, finally, designing a clear authorship policy and ensuring the appropriate publication of the study results. We hope that these suggestions encourage potential researchers to conduct multicenter studies, thereby collectively enhancing the quality of research and its application to clinical practice.

There is no such thing as a validated prediction model

BACKGROUND

Clinical prediction models should be validated before implementation in clinical practice. But is favorable performance at internal validation or one external validation sufficient to claim that a prediction model works well in the intended clinical context?

MAIN BODY

We argue to the contrary because (1) patient populations vary, (2) measurement procedures vary, and (3) populations and measurements change over time. Hence, we have to expect heterogeneity in model performance between locations and settings, and across time. It follows that prediction models are never truly validated. This does not imply that validation is not important. Rather, the current focus on developing new models should shift to a focus on more extensive, well-conducted, and well-reported validation studies of promising models.

CONCLUSION

Principled validation strategies are needed to understand and quantify heterogeneity, monitor performance over time, and update prediction models when appropriate. Such strategies will help to ensure that prediction models stay up-to-date and safe to support clinical decision-making.

Reporting of Model Performance and Statistical Methods in Studies That Use Machine Learning to Develop Clinical Prediction Models: Protocol for a Systematic Review

BACKGROUND

With the growing excitement of the potential benefits of using machine learning and artificial intelligence in medicine, the number of published clinical prediction models that use these approaches has increased. However, there is evidence (albeit limited) that suggests that the reporting of machine learning-specific aspects in these studies is poor. Further, there are no reviews assessing the reporting quality or broadly accepted reporting guidelines for these aspects.

OBJECTIVE

This paper presents the protocol for a systematic review that will assess the reporting quality of machine learning-specific aspects in studies that use machine learning to develop clinical prediction models.

METHODS

We will include studies that use a supervised machine learning algorithm to develop a prediction model for use in clinical practice (ie, for diagnosis or prognosis of a condition or identification of candidates for health care interventions). We will search MEDLINE for studies published in 2019, pseudorandomly sort the records, and screen until we obtain 100 studies that meet our inclusion criteria. We will assess reporting quality with a novel checklist developed in parallel with this review, which includes content derived from existing reporting guidelines, textbooks, and consultations with experts. The checklist will cover 4 key areas where the reporting of machine learning studies is unique: modelling steps (order and data used for each step), model performance (eg, reporting the performance of each model compared), statistical methods (eg, describing the tuning approach), and presentation of models (eg, specifying the predictors that contributed to the final model).

RESULTS

We completed data analysis in August 2021 and are writing the manuscript. We expect to submit the results to a peer-reviewed journal in early 2022.

CONCLUSIONS

This review will contribute to more standardized and complete reporting in the field by identifying areas where reporting is poor and can be improved.

TRIAL REGISTRATION

PROSPERO International Prospective Register of Systematic Reviews CRD42020206167; https://www.crd.york.ac.uk/PROSPERO/display_record.php?RecordID=206167.

INTERNATIONAL REGISTERED REPORT IDENTIFIER (IRRID)

RR1-10.2196/30956.

Machine Learning for Clinical Decision-Making: Challenges and Opportunities in Cardiovascular Imaging

The use of machine learning (ML) approaches to target clinical problems is called to revolutionize clinical decision-making in cardiology. The success of these tools is dependent on the understanding of the intrinsic processes being used during the conventional pathway by which clinicians make decisions. In a parallelism with this pathway, ML can have an impact at four levels: for data acquisition, predominantly by extracting standardized, high-quality information with the smallest possible learning curve; for feature extraction, by discharging healthcare practitioners from performing tedious measurements on raw data; for interpretation, by digesting complex, heterogeneous data in order to augment the understanding of the patient status; and for decision support, by leveraging the previous steps to predict clinical outcomes, response to treatment or to recommend a specific intervention. This paper discusses the state-of-the-art, as well as the current clinical status and challenges associated with the two later tasks of interpretation and decision support, together with the challenges related to the learning process, the auditability/traceability, the system infrastructure and the integration within clinical processes in cardiovascular imaging.

Risk prediction in multicentre studies when there is confounding by cluster or informative cluster size

Background

Clustered data arise in research when patients are clustered within larger units. Generalised Estimating Equations (GEE) and Generalised Linear Models (GLMM) can be used to provide marginal and cluster-specific inference and predictions, respectively.

Methods

Confounding by Cluster (CBC) and Informative cluster size (ICS) are two complications that may arise when modelling clustered data. CBC can arise when the distribution of a predictor variable (termed ‘exposure’), varies between clusters causing confounding of the exposure-outcome relationship. ICS means that the cluster size conditional on covariates is not independent of the outcome. In both situations, standard GEE and GLMM may provide biased or misleading inference, and modifications have been proposed. However, both CBC and ICS are routinely overlooked in the context of risk prediction, and their impact on the predictive ability of the models has been little explored. We study the effect of CBC and ICS on the predictive ability of risk models for binary outcomes when GEE and GLMM are used. We examine whether two simple approaches to handle CBC and ICS, which involve adjusting for the cluster mean of the exposure and the cluster size, respectively, can improve the accuracy of predictions.

Results

Both CBC and ICS can be viewed as violations of the assumptions in the standard GLMM; the random effects are correlated with exposure for CBC and cluster size for ICS. Based on these principles, we simulated data subject to CBC/ICS. The simulation studies suggested that the predictive ability of models derived from using standard GLMM and GEE ignoring CBC/ICS was affected. Marginal predictions were found to be mis-calibrated. Adjusting for the cluster-mean of the exposure or the cluster size improved calibration, discrimination and the overall predictive accuracy of marginal predictions, by explaining part of the between cluster variability. The presence of CBC/ICS did not affect the accuracy of conditional predictions. We illustrate these concepts using real data from a multicentre study with potential CBC.

Conclusion

Ignoring CBC and ICS when developing prediction models for clustered data can affect the accuracy of marginal predictions. Adjusting for the cluster mean of the exposure or the cluster size can improve the predictive accuracy of marginal predictions.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12874-021-01321-x.

Clinical decision support systems in orthodontics: A narrative review of data science approaches

Advancements in technology and data collection generated immense amounts of information from various sources such as health records, clinical examination, imaging, medical devices, as well as experimental and biological data. Proper management and analysis of these data via high-end computing solutions, artificial intelligence and machine learning approaches can assist in extracting meaningful information that enhances population health and well-being. Furthermore, the extracted knowledge can provide new avenues for modern healthcare delivery via clinical decision support systems. This manuscript presents a narrative review of data science approaches for clinical decision support systems in orthodontics. We describe the fundamental components of data science approaches including (a) Data collection, storage and management; (b) Data processing; (c) In-depth data analysis; and (d) Data communication. Then, we introduce a web-based data management platform, the Data Storage for Computation and Integration, for temporomandibular joint and dental clinical decision support systems.

Development and validation of a clinical prediction rule for development of diabetic foot ulceration: an analysis of data from five cohort studies

INTRODUCTION

The aim of the study was to develop and validate a clinical prediction rule (CPR) for foot ulceration in people with diabetes.

RESEARCH DESIGN AND METHODS

Development of a CPR using individual participant data from four international cohort studies identified by systematic review, with validation in a fifth study. Development cohorts were from primary and secondary care foot clinics in Europe and the USA (n=8255, adults over 18 years old, with diabetes, ulcer free at recruitment). Using data from monofilament testing, presence/absence of pulses, and participant history of previous ulcer and/or amputation, we developed a simple CPR to predict who will develop a foot ulcer within 2 years of initial assessment and validated it in a fifth study (n=3324). The CPR's performance was assessed with C-statistics, calibration slopes, calibration-in-the-large, and a net benefit analysis.

RESULTS

CPR scores of 0, 1, 2, 3, and 4 had a risk of ulcer within 2 years of 2.4% (95% CI 1.5% to 3.9%), 6.0% (95% CI 3.5% to 9.5%), 14.0% (95% CI 8.5% to 21.3%), 29.2% (95% CI 19.2% to 41.0%), and 51.1% (95% CI 37.9% to 64.1%), respectively. In the validation dataset, calibration-in-the-large was -0.374 (95% CI -0.561 to -0.187) and calibration slope 1.139 (95% CI 0.994 to 1.283). The C-statistic was 0.829 (95% CI 0.790 to 0.868). The net benefit analysis suggested that people with a CPR score of 1 or more (risk of ulceration 6.0% or more) should be referred for treatment.

CONCLUSION

The clinical prediction rule is simple, using routinely obtained data, and could help prevent foot ulcers by redirecting care to patients with scores of 1 or above. It has been validated in a community setting, and requires further validation in secondary care settings.

The myth of generalisability in clinical research and machine learning in health care

An emphasis on overly broad notions of generalisability as it pertains to applications of machine learning in health care can overlook situations in which machine learning might provide clinical utility. We believe that this narrow focus on generalisability should be replaced with wider considerations for the ultimate goal of building machine learning systems that are useful at the bedside.

Early warning scores for detecting deterioration in adult hospital patients: systematic review and critical appraisal of methodology

OBJECTIVE

To provide an overview and critical appraisal of early warning scores for adult hospital patients.

DESIGN

Systematic review.

DATA SOURCES

Medline, CINAHL, PsycInfo, and Embase until June 2019.

ELIGIBILITY CRITERIA FOR STUDY SELECTION

Studies describing the development or external validation of an early warning score for adult hospital inpatients.

RESULTS

13 171 references were screened and 95 articles were included in the review. 11 studies were development only, 23 were development and external validation, and 61 were external validation only. Most early warning scores were developed for use in the United States (n=13/34, 38%) and the United Kingdom (n=10/34, 29%). Death was the most frequent prediction outcome for development studies (n=10/23, 44%) and validation studies (n=66/84, 79%), with different time horizons (the most frequent was 24 hours). The most common predictors were respiratory rate (n=30/34, 88%), heart rate (n=28/34, 83%), oxygen saturation, temperature, and systolic blood pressure (all n=24/34, 71%). Age (n=13/34, 38%) and sex (n=3/34, 9%) were less frequently included. Key details of the analysis populations were often not reported in development studies (n=12/29, 41%) or validation studies (n=33/84, 39%). Small sample sizes and insufficient numbers of event patients were common in model development and external validation studies. Missing data were often discarded, with just one study using multiple imputation. Only nine of the early warning scores that were developed were presented in sufficient detail to allow individualised risk prediction. Internal validation was carried out in 19 studies, but recommended approaches such as bootstrapping or cross validation were rarely used (n=4/19, 22%). Model performance was frequently assessed using discrimination (development n=18/22, 82%; validation n=69/84, 82%), while calibration was seldom assessed (validation n=13/84, 15%). All included studies were rated at high risk of bias.

CONCLUSIONS

Early warning scores are widely used prediction models that are often mandated in daily clinical practice to identify early clinical deterioration in hospital patients. However, many early warning scores in clinical use were found to have methodological weaknesses. Early warning scores might not perform as well as expected and therefore they could have a detrimental effect on patient care. Future work should focus on following recommended approaches for developing and evaluating early warning scores, and investigating the impact and safety of using these scores in clinical practice.

SYSTEMATIC REVIEW REGISTRATION

PROSPERO CRD42017053324.

Assessment of an Updated Neonatal Research Network Extremely Preterm Birth Outcome Model in the Vermont Oxford Network

IMPORTANCE

The Eunice Kennedy Shriver National Institute of Child Health and Human Development Neonatal Research Network (NRN) extremely preterm birth outcome model is widely used for prognostication by practitioners caring for families expecting extremely preterm birth. The model provides information on mean outcomes from 1998 to 2003 and does not account for substantial variation in outcomes among US hospitals.

OBJECTIVE

To update and validate the NRN extremely preterm birth outcome model for most extremely preterm infants in the United States.

DESIGN, SETTING, AND PARTICIPANTS

This prognostic study included 3 observational cohorts from January 1, 2006, to December 31, 2016, at 19 US centers in the NRN (derivation cohort) and 637 US centers in Vermont Oxford Network (VON) (validation cohorts). Actively treated infants born at 22 weeks' 0 days' to 25 weeks' 6 days' gestation and weighing 401 to 1000 g, including 4176 in the NRN for 2006 to 2012, 45 179 in VON for 2006 to 2012, and 25 969 in VON for 2013 to 2016, were studied. VON cohorts comprised more than 85% of eligible US births. Data analysis was performed from May 1, 2017, to March 31, 2019.

EXPOSURES

Predictive variables used in the original model, including infant sex, birth weight, plurality, gestational age at birth, and exposure to antenatal corticosteroids.

MAIN OUTCOMES AND MEASURES

The main outcome was death before discharge. Secondary outcomes included neurodevelopmental impairment at 18 to 26 months' corrected age and measures of hospital resource use (days of hospitalization and ventilator use).

RESULTS

Among 4176 actively treated infants in the NRN cohort (48% female; mean [SD] gestational age, 24.2 [0.8] weeks), survival was 63% vs 62% among 3702 infants in the era of the original model (47% female; mean [SD] gestational age, 24.2 [0.8] weeks). In the concurrent (2006-2012) VON cohort, survival was 66% among 45 179 actively treated infants (47% female; mean [SD] gestational age, 24.1 [0.8] weeks) and 70% among 25 969 infants from 2013 to 2016 (48% female; mean [SD] gestational age, 24.1 [0.8] weeks). Model C statistics were 0.74 in the 2006-2012 validation cohort and 0.73 in the 2013-2016 validation cohort. With the use of decision curve analysis to compare the model with a gestational age-only approach to prognostication, the updated model showed a predictive advantage. The birth hospital contributed equally as much to prediction of survival as gestational age (20%) but less than the other factors combined (60%).

CONCLUSIONS AND RELEVANCE

An updated model using well-known factors to predict survival for extremely preterm infants performed moderately well when applied to large US cohorts. Because survival rates change over time, the model requires periodic updating. The hospital of birth contributed substantially to outcome prediction.

Prediction models for diagnosis and prognosis of covid-19: systematic review and critical appraisal

OBJECTIVE

To review and appraise the validity and usefulness of published and preprint reports of prediction models for diagnosing coronavirus disease 2019 (covid-19) in patients with suspected infection, for prognosis of patients with covid-19, and for detecting people in the general population at increased risk of covid-19 infection or being admitted to hospital with the disease.

DESIGN

Living systematic review and critical appraisal by the COVID-PRECISE (Precise Risk Estimation to optimise covid-19 Care for Infected or Suspected patients in diverse sEttings) group.

DATA SOURCES

PubMed and Embase through Ovid, up to 1 July 2020, supplemented with arXiv, medRxiv, and bioRxiv up to 5 May 2020.

STUDY SELECTION

Studies that developed or validated a multivariable covid-19 related prediction model.

DATA EXTRACTION

At least two authors independently extracted data using the CHARMS (critical appraisal and data extraction for systematic reviews of prediction modelling studies) checklist; risk of bias was assessed using PROBAST (prediction model risk of bias assessment tool).

RESULTS

37 421 titles were screened, and 169 studies describing 232 prediction models were included. The review identified seven models for identifying people at risk in the general population; 118 diagnostic models for detecting covid-19 (75 were based on medical imaging, 10 to diagnose disease severity); and 107 prognostic models for predicting mortality risk, progression to severe disease, intensive care unit admission, ventilation, intubation, or length of hospital stay. The most frequent types of predictors included in the covid-19 prediction models are vital signs, age, comorbidities, and image features. Flu-like symptoms are frequently predictive in diagnostic models, while sex, C reactive protein, and lymphocyte counts are frequent prognostic factors. Reported C index estimates from the strongest form of validation available per model ranged from 0.71 to 0.99 in prediction models for the general population, from 0.65 to more than 0.99 in diagnostic models, and from 0.54 to 0.99 in prognostic models. All models were rated at high or unclear risk of bias, mostly because of non-representative selection of control patients, exclusion of patients who had not experienced the event of interest by the end of the study, high risk of model overfitting, and unclear reporting. Many models did not include a description of the target population (n=27, 12%) or care setting (n=75, 32%), and only 11 (5%) were externally validated by a calibration plot. The Jehi diagnostic model and the 4C mortality score were identified as promising models.

CONCLUSION

Prediction models for covid-19 are quickly entering the academic literature to support medical decision making at a time when they are urgently needed. This review indicates that almost all pubished prediction models are poorly reported, and at high risk of bias such that their reported predictive performance is probably optimistic. However, we have identified two (one diagnostic and one prognostic) promising models that should soon be validated in multiple cohorts, preferably through collaborative efforts and data sharing to also allow an investigation of the stability and heterogeneity in their performance across populations and settings. Details on all reviewed models are publicly available at https://www.covprecise.org/. Methodological guidance as provided in this paper should be followed because unreliable predictions could cause more harm than benefit in guiding clinical decisions. Finally, prediction model authors should adhere to the TRIPOD (transparent reporting of a multivariable prediction model for individual prognosis or diagnosis) reporting guideline.

SYSTEMATIC REVIEW REGISTRATION

Protocol https://osf.io/ehc47/, registration https://osf.io/wy245.

READERS' NOTE

This article is a living systematic review that will be updated to reflect emerging evidence. Updates may occur for up to two years from the date of original publication. This version is update 3 of the original article published on 7 April 2020 (BMJ 2020;369:m1328). Previous updates can be found as data supplements (https://www.bmj.com/content/369/bmj.m1328/related#datasupp). When citing this paper please consider adding the update number and date of access for clarity.

Developing risk models for multicenter data using standard logistic regression produced suboptimal predictions: A simulation study

Although multicenter data are common, many prediction model studies ignore this during model development. The objective of this study is to evaluate the predictive performance of regression methods for developing clinical risk prediction models using multicenter data, and provide guidelines for practice. We compared the predictive performance of standard logistic regression, generalized estimating equations, random intercept logistic regression, and fixed effects logistic regression. First, we presented a case study on the diagnosis of ovarian cancer. Subsequently, a simulation study investigated the performance of the different models as a function of the amount of clustering, development sample size, distribution of center‐specific intercepts, the presence of a center‐predictor interaction, and the presence of a dependency between center effects and predictors. The results showed that when sample sizes were sufficiently large, conditional models yielded calibrated predictions, whereas marginal models yielded miscalibrated predictions. Small sample sizes led to overfitting and unreliable predictions. This miscalibration was worse with more heavily clustered data. Calibration of random intercept logistic regression was better than that of standard logistic regression even when center‐specific intercepts were not normally distributed, a center‐predictor interaction was present, center effects and predictors were dependent, or when the model was applied in a new center. Therefore, to make reliable predictions in a specific center, we recommend random intercept logistic regression.

Evidence synthesis in prognosis research

Over the past few years, evidence synthesis has become essential to investigate and improve the generalizability of medical research findings. This strategy often involves a meta-analysis to formally summarize quantities of interest, such as relative treatment effect estimates. The use of meta-analysis methods is, however, less straightforward in prognosis research because substantial variation exists in research objectives, analysis methods and the level of reported evidence. We present a gentle overview of statistical methods that can be used to summarize data of prognostic factor and prognostic model studies. We discuss how aggregate data, individual participant data, or a combination thereof can be combined through meta-analysis methods. Recent examples are provided throughout to illustrate the various methods.