Identifying clinical features and blood biomarkers associated with mild cognitive impairment in Parkinson disease using machine learning.

BACKGROUND AND PURPOSE: A broad list of variables associated with mild cognitive impairment (MCI) in Parkinson disease (PD) have been investigated separately. However, there is as yet no study including all of them to assess variable importance. Shapley variable importance cloud (ShapleyVIC) can robustly assess variable importance while accounting for correlation between variables. Objectives of this study were (i) to prioritize the important variables associated with PD-MCI and (ii) to explore new blood biomarkers related to PD-MCI. METHODS: ShapleyVIC-assisted variable selection was used to identify a subset of variables from 41 variables potentially associated with PD-MCI in a cross-sectional study. Backward selection was used to further identify the variables associated with PD-MCI. Relative risk was used to quantify the association of final associated variables and PD-MCI in the final multivariable log-binomial regression model. RESULTS: Among 41 variables analysed, 22 variables were identified as significantly important variables associated with PD-MCI and eight variables were subsequently selected in the final model, indicating fewer years of education, shorter history of hypertension, higher Movement Disorder Society-Unified Parkinson's Disease Rating Scale motor score, higher levels of triglyceride (TG) and apolipoprotein A1 (ApoA1), and SNCA rs6826785 noncarrier status were associated with increased risk of PD-MCI (p < 0.05). CONCLUSIONS: Our study highlighted the strong association between TG, ApoA1, SNCA rs6826785, and PD-MCI by machine learning approach. Screening and management of high TG and ApoA1 levels might help prevent cognitive impairment in early PD patients. SNCA rs6826785 could be a novel therapeutic target for PD-MCI. ShapleyVIC-assisted variable selection is a novel and robust alternative to traditional approaches for future clinical study to prioritize the variables of interest.

FedScore: A privacy-preserving framework for federated scoring system development.

OBJECTIVE: We propose FedScore, a privacy-preserving federated learning framework for scoring system generation across multiple sites to facilitate cross-institutional collaborations. MATERIALS AND METHODS: The FedScore framework includes five modules: federated variable ranking, federated variable transformation, federated score derivation, federated model selection and federated model evaluation. To illustrate usage and assess FedScore's performance, we built a hypothetical global scoring system for mortality prediction within 30 days after a visit to an emergency department using 10 simulated sites divided from a tertiary hospital in Singapore. We employed a pre-existing score generator to construct 10 local scoring systems independently at each site and we also developed a scoring system using centralized data for comparison. RESULTS: We compared the acquired FedScore model's performance with that of other scoring models using the receiver operating characteristic (ROC) analysis. The FedScore model achieved an average area under the curve (AUC) value of 0.763 across all sites, with a standard deviation (SD) of 0.020. We also calculated the average AUC values and SDs for each local model, and the FedScore model showed promising accuracy and stability with a high average AUC value which was closest to the one of the pooled model and SD which was lower than that of most local models. CONCLUSION: This study demonstrates that FedScore is a privacy-preserving scoring system generator with potentially good generalizability.

Estimating risk ratio from any standard epidemiological design by doubling the cases.

BACKGROUND: Despite the ease of interpretation and communication of a risk ratio (RR), and several other advantages in specific settings, the odds ratio (OR) is more commonly reported in epidemiological and clinical research. This is due to the familiarity of the logistic regression model for estimating adjusted ORs from data gathered in a cross-sectional, cohort or case-control design. The preservation of the OR (but not RR) in case-control samples has contributed to the perception that it is the only valid measure of relative risk from case-control samples. For cohort or cross-sectional data, a method known as 'doubling-the-cases' provides valid estimates of RR and an expression for a robust standard error has been derived, but is not available in statistical software packages. METHODS: In this paper, we first describe the doubling-of-cases approach in the cohort setting and then extend its application to case-control studies by incorporating sampling weights and deriving an expression for a robust standard error. The performance of the estimator is evaluated using simulated data, and its application illustrated in a study of neonatal jaundice. We provide an R package that implements the method for any standard design. RESULTS: Our work illustrates that the doubling-of-cases approach for estimating an adjusted RR from cross-sectional or cohort data can also yield valid RR estimates from case-control data. The approach is straightforward to apply, involving simple modification of the data followed by logistic regression analysis. The method performed well for case-control data from simulated cohorts with a range of prevalence rates. In the application to neonatal jaundice, the RR estimates were similar to those from relative risk regression, whereas the OR from naive logistic regression overestimated the RR despite the low prevalence of the outcome. CONCLUSIONS: By providing an R package that estimates an adjusted RR from cohort, cross-sectional or case-control studies, we have enabled the method to be easily implemented with familiar software, so that investigators are not limited to reporting an OR and can examine the RR when it is of interest.

AutoScore-Ordinal: an interpretable machine learning framework for generating scoring models for ordinal outcomes.

BACKGROUND: Risk prediction models are useful tools in clinical decision-making which help with risk stratification and resource allocations and may lead to a better health care for patients. AutoScore is a machine learning-based automatic clinical score generator for binary outcomes. This study aims to expand the AutoScore framework to provide a tool for interpretable risk prediction for ordinal outcomes. METHODS: The AutoScore-Ordinal framework is generated using the same 6 modules of the original AutoScore algorithm including variable ranking, variable transformation, score derivation (from proportional odds models), model selection, score fine-tuning, and model evaluation. To illustrate the AutoScore-Ordinal performance, the method was conducted on electronic health records data from the emergency department at Singapore General Hospital over 2008 to 2017. The model was trained on 70% of the data, validated on 10% and tested on the remaining 20%. RESULTS: This study included 445,989 inpatient cases, where the distribution of the ordinal outcome was 80.7% alive without 30-day readmission, 12.5% alive with 30-day readmission, and 6.8% died inpatient or by day 30 post discharge. Two point-based risk prediction models were developed using two sets of 8 predictor variables identified by the flexible variable selection procedure. The two models indicated reasonably good performance measured by mean area under the receiver operating characteristic curve (0.758 and 0.793) and generalized c-index (0.737 and 0.760), which were comparable to alternative models. CONCLUSION: AutoScore-Ordinal provides an automated and easy-to-use framework for development and validation of risk prediction models for ordinal outcomes, which can systematically identify potential predictors from high-dimensional data.

AutoScore-Survival: Developing interpretable machine learning-based time-to-event scores with right-censored survival data.

BACKGROUND: Scoring systems are highly interpretable and widely used to evaluate time-to-event outcomes in healthcare research. However, existing time-to-event scores are predominantly created ad-hoc using a few manually selected variables based on clinician's knowledge, suggesting an unmet need for a robust and efficient generic score-generating method. METHODS: AutoScore was previously developed as an interpretable machine learning score generator, integrating both machine learning and point-based scores in the strong discriminability and accessibility. We have further extended it to the time-to-event outcomes and developed AutoScore-Survival, for generating time-to-event scores with right-censored survival data. Random survival forest provided an efficient solution for selecting variables, and Cox regression was used for score weighting. We implemented our proposed method as an R package. We illustrated our method in a study of 90-day survival prediction for patients in intensive care units and compared its performance with other survival models, the random survival forest, and two traditional clinical scores. RESULTS: The AutoScore-Survival-derived scoring system was more parsimonious than survival models built using traditional variable selection methods (e.g., penalized likelihood approach and stepwise variable selection), and its performance was comparable to survival models using the same set of variables. Although AutoScore-Survival achieved a comparable integrated area under the curve of 0.782 (95% CI: 0.767-0.794), the integer-valued time-to-event scores generated are favorable in clinical applications because they are easier to compute and interpret. CONCLUSIONS: Our proposed AutoScore-Survival provides a robust and easy-to-use machine learning-based clinical score generator to studies of time-to-event outcomes. It gives a systematic guideline to facilitate the future development of time-to-event scores for clinical applications.

Deep learning for temporal data representation in electronic health records: A systematic review of challenges and methodologies.

OBJECTIVE: Temporal electronic health records (EHRs) contain a wealth of information for secondary uses, such as clinical events prediction and chronic disease management. However, challenges exist for temporal data representation. We therefore sought to identify these challenges and evaluate novel methodologies for addressing them through a systematic examination of deep learning solutions. METHODS: We searched five databases (PubMed, Embase, the Institute of Electrical and Electronics Engineers [IEEE] Xplore Digital Library, the Association for Computing Machinery [ACM] Digital Library, and Web of Science) complemented with hand-searching in several prestigious computer science conference proceedings. We sought articles that reported deep learning methodologies on temporal data representation in structured EHR data from January 1, 2010, to August 30, 2020. We summarized and analyzed the selected articles from three perspectives: nature of time series, methodology, and model implementation. RESULTS: We included 98 articles related to temporal data representation using deep learning. Four major challenges were identified, including data irregularity, heterogeneity, sparsity, and model opacity. We then studied how deep learning techniques were applied to address these challenges. Finally, we discuss some open challenges arising from deep learning. CONCLUSION: Temporal EHR data present several major challenges for clinical prediction modeling and data utilization. To some extent, current deep learning solutions can address these challenges. Future studies may consider designing comprehensive and integrated solutions. Moreover, researchers should incorporate clinical domain knowledge into study designs and enhance model interpretability to facilitate clinical implementation.

AutoScore-Imbalance: An interpretable machine learning tool for development of clinical scores with rare events data.

BACKGROUND: Medical decision-making impacts both individual and public health. Clinical scores are commonly used among various decision-making models to determine the degree of disease deterioration at the bedside. AutoScore was proposed as a useful clinical score generator based on machine learning and a generalized linear model. However, its current framework still leaves room for improvement when addressing unbalanced data of rare events. METHODS: Using machine intelligence approaches, we developed AutoScore-Imbalance, which comprises three components: training dataset optimization, sample weight optimization, and adjusted AutoScore. Baseline techniques for performance comparison included the original AutoScore, full logistic regression, stepwise logistic regression, least absolute shrinkage and selection operator (LASSO), full random forest, and random forest with a reduced number of variables. These models were evaluated based on their area under the curve (AUC) in the receiver operating characteristic analysis and balanced accuracy (i.e., mean value of sensitivity and specificity). By utilizing a publicly accessible dataset from Beth Israel Deaconess Medical Center, we assessed the proposed model and baseline approaches to predict inpatient mortality. RESULTS: AutoScore-Imbalance outperformed baselines in terms of AUC and balanced accuracy. The nine-variable AutoScore-Imbalance sub-model achieved the highest AUC of 0.786 (0.732-0.839), while the eleven-variable original AutoScore obtained an AUC of 0.723 (0.663-0.783), and the logistic regression with 21 variables obtained an AUC of 0.743 (0.685-0.801). The AutoScore-Imbalance sub-model (using a down-sampling algorithm) yielded an AUC of 0.771 (0.718-0.823) with only five variables, demonstrating a good balance between performance and variable sparsity. Furthermore, AutoScore-Imbalance obtained the highest balanced accuracy of 0.757 (0.702-0.805), compared to 0.698 (0.643-0.753) by the original AutoScore and the maximum of 0.720 (0.664-0.769) by other baseline models. CONCLUSIONS: We have developed an interpretable tool to handle clinical data imbalance, presented its structure, and demonstrated its superiority over baselines. The AutoScore-Imbalance tool can be applied to highly unbalanced datasets to gain further insight into rare medical events and facilitate real-world clinical decision-making.

Robust estimation of the effect of an exposure on the change in a continuous outcome.

BACKGROUND: The change in two measurements of a continuous outcome can be modelled directly with a linear regression model, or indirectly with a random effects model (REM) of the individual measurements. These methods are susceptible to model misspecifications, which are commonly addressed by applying monotonic transformations (e.g., Box-Cox transformation) to the outcomes. However, transforming the outcomes complicates the data analysis, especially when variable selection is involved. We propose a robust alternative through a novel application of the conditional probit (cprobit) model. METHODS: The cprobit model analyzes the ordered outcomes within each subject, making the estimate invariant to monotonic transformation on the outcome. By scaling the estimate from the cprobit model, we obtain the exposure effect on the change in the observed or Box-Cox transformed outcome, pending the adequacy of the normality assumption on the raw or transformed scale. RESULTS: Using simulated data, we demonstrated a similar good performance of the cprobit model and REM with and without transformation, except for some bias from both methods when the Box-Cox transformation was applied to scenarios with small sample size and strong effects. Only the cprobit model was robust to skewed subject-specific intercept terms when a Box-Cox transformation was used. Using two real datasets from the breast cancer and inpatient glycemic variability studies which utilize electronic medical records, we illustrated the application of our proposed robust approach as a seamless three-step workflow that facilitates the use of Box-Cox transformation to address non-normality with a common underlying model. CONCLUSIONS: The cprobit model provides a seamless and robust inference on the change in continuous outcomes, and its three-step workflow is implemented in an R package for easy accessibility.

Using marginal standardisation to estimate relative risk without dichotomising continuous outcomes.

BACKGROUND: Although criticisms regarding the dichotomisation of continuous variables are well known, applying logit model to dichotomised outcomes is the convention because the odds ratios are easily obtained and they approximate the relative risks (RRs) for rare events. METHODS: To avoid dichotomisation when estimating RR, the marginal standardisation method that transforms estimates from logit or probit model to RR estimate is extended to include estimates from linear model in the transformation. We conducted a simulation study to compare the statistical properties of the estimates from: (i) marginal standardisation method between models for continuous (i.e., linear model) and dichotomised outcomes (i.e., logit or probit model), and (ii) marginal standardisation method and distributional approach (i.e., marginal mean method) applied to linear model. We also compared the diagnostic test for probit, logit and linear models. For the real dataset analysis, we applied these analytical approaches to assess the management of inpatient hyperglycaemia in a pilot intervention study. RESULTS: Although the RR estimates from the marginal standardisation method were generally unbiased for all models in the simulation study, the marginal standardisation method for linear model provided estimates with higher precision and power than logit or probit model, especially when the baseline risks were at the extremes. When comparing approaches that avoid dichotomisation, RR estimates from these approaches had comparable performance. Assessing the assumption of error distribution was less powerful for logit or probit model via link test when compared with diagnostic test for linear model. After accounting for multiple thresholds representing varying levels of severity in hyperglycaemia, marginal standardisation method for linear model provided stronger evidence of reduced hyperglycaemia risk after intervention in the real dataset analysis although the RR estimates were similar across various approaches. CONCLUSIONS: When compared with approaches that do not avoid dichotomisation, the RR estimated from linear model is more precise and powerful, and the diagnostic test from linear model is more powerful in detecting mis-specified error distributional assumption than the diagnostic test from logit or probit model. Our work describes and assesses the methods available to analyse data involving studies of continuous outcomes with binary representations.

Individual and Population Trajectories of Influenza Antibody Titers Over Multiple Seasons in a Tropical Country.

Seasonal influenza epidemics occur year-round in the tropics, complicating the planning of vaccination programs. We built an individual-level longitudinal model of baseline antibody levels, time of infection, and the subsequent rise and decay of antibodies postinfection using influenza A(H1N1)pdm09 data from 2 sources in Singapore: 1) a noncommunity cohort with real-time polymerase chain reaction-confirmed infections and at least 1 serological sample collected from each participant between May and October 2009 (n = 118) and 2) a community cohort with up to 6 serological samples collected between May 2009 and October 2010 (n = 760). The model was hierarchical, to account for interval censoring and interindividual variation. Model parameters were estimated via a reversible jump Markov chain Monte Carlo algorithm using custom-designed R (https://www.r-project.org/) and C++ (https://isocpp.org/) code. After infection, antibody levels peaked at 4-7 weeks, with a half-life of 26.5 weeks, followed by a slower decrease up to 1 year to approximately preinfection levels. After the third wave, the seropositivity rate and the population-level antibody titer dropped to the same level as they were at the end of the first pandemic wave. The results of this analysis are consistent with the hypothesis that the population-level effect of individuals' waxing and waning antibodies influences influenza seasonality in the tropics.

Quantifying temporal trends of age-standardized rates with odds.

BACKGROUND: To quantify temporal trends in age-standardized rates of disease, the convention is to fit a linear regression model to log-transformed rates because the slope term provides the estimated annual percentage change. However, such log-transformation is not always appropriate. METHODS: We propose an alternative method using the rank-ordered logit (ROL) model that is indifferent to log-transformation. This method quantifies the temporal trend using odds, a quantity commonly used in epidemiology, and the log-odds corresponds to the scaled slope parameter estimate from linear regression. The ROL method can be implemented by using the commands for proportional hazards regression in any standard statistical package. We apply the ROL method to estimate temporal trends in age-standardized cancer rates worldwide using the cancer incidence data from the Cancer Incidence in Five Continents plus (CI5plus) database for the period 1953 to 2007 and compare the estimates to their scaled counterparts obtained from linear regression with and without log-transformation. RESULTS: We found a strong concordance in the direction and significance of the temporal trends in cancer incidence estimated by all three approaches, and illustrated how the estimate from the ROL model provides a measure that is comparable to a scaled slope parameter estimated from linear regression. CONCLUSIONS: Our method offers an alternative approach for quantifying temporal trends in incidence or mortality rates in a population that is invariant to transformation, and whose estimate of trend agrees with the scaled slope from a linear regression model.

Utilizing distributional analytics and electronic records to assess timeliness of inpatient blood glucose monitoring in non-critical care wards.

BACKGROUND: Regular and timely monitoring of blood glucose (BG) levels in hospitalized patients with diabetes mellitus is crucial to optimizing inpatient glycaemic control. However, methods to quantify timeliness as a measurement of quality of care are lacking. We propose an analytical approach that utilizes BG measurements from electronic records to assess adherence to an inpatient BG monitoring protocol in hospital wards. METHODS: We applied our proposed analytical approach to electronic records obtained from 24 non-critical care wards in November and December 2013 from a tertiary care hospital in Singapore. We applied distributional analytics to evaluate daily adherence to BG monitoring timings. A one-sample Kolmogorov-Smirnov (1S-KS) test was performed to test daily BG timings against non-adherence represented by the uniform distribution. This test was performed among wards with high power, determined through simulation. The 1S-KS test was coupled with visualization via the cumulative distribution function (cdf) plot and a two-sample Kolmogorov-Smirnov (2S-KS) test, enabling comparison of the BG timing distributions between two consecutive days. We also applied mixture modelling to identify the key features in daily BG timings. RESULTS: We found that 11 out of the 24 wards had high power. Among these wards, 1S-KS test with cdf plots indicated adherence to BG monitoring protocols. Integrating both 1S-KS and 2S-KS information within a moving window consisting of two consecutive days did not suggest frequent potential change from or towards non-adherence to protocol. From mixture modelling among wards with high power, we consistently identified four components with high concentration of BG measurements taken before mealtimes and around bedtime. This agnostic analysis provided additional evidence that the wards were adherent to BG monitoring protocols. CONCLUSIONS: We demonstrated the utility of our proposed analytical approach as a monitoring tool. It provided information to healthcare providers regarding the timeliness of daily BG measurements. From the real data application, there were empirical evidences suggesting adherence of BG timings to protocol among wards with adequate power for assessing timeliness. Our approach is extendable to other areas of healthcare where timeliness of patient care processes is important.

Predicting Satisfaction for Unicompartmental Knee Arthroplasty Patients in an Asian Population.

BACKGROUND: Despite renewed interest in unicompartmental knee arthroplasty (UKA), there is a paucity of published literature with regard to patient satisfaction after UKA within Asian populations. The purpose of this study is to identify characteristics and factors which may contribute to patient dissatisfaction after UKA in a multiracial Asian population. METHODS: Seven hundred twenty-four UKAs were performed between January 2007 and April 2013. Preoperative and postoperative variables were prospectively captured, such as standardized knee scores, knee range of motion, and patient satisfaction scores. These variables were then analyzed with a multiple logistic regression model to determine statistically significant factors contributing to patients' satisfaction. RESULTS: Minimum duration of follow-up was 2 years, with an overall patient satisfaction rate of 92.2%. There was improvement in mean knee range of motion and across various standardized knee scores. Preoperative variables associated with patient dissatisfaction included a poorer preoperative Mental Component Summary, better preoperative knee extension, and better preoperative Oxford Knee Scores. Significant postoperative variables included better Oxford Knee Score at 6 months and Mental Component Summary at 2 years. CONCLUSION: Despite the impressive patient satisfaction rate of UKA in this Asian population, these findings suggest that there is a targeted group of patients with select preoperative factors who would benefit from preoperative counseling.

Topical Atropine for Childhood Myopia Control: The Atropine Treatment Long-Term Assessment Study.

Importance: Clinical trial results of topical atropine eye drops for childhood myopia control have shown inconsistent outcomes across short-term studies, with little long-term safety or other outcomes reported. Objective: To report the long-term safety and outcomes of topical atropine for childhood myopia control. Design, Setting, and Participants: This prospective, double-masked observational study of the Atropine for the Treatment of Myopia (ATOM) 1 and ATOM2 randomized clinical trials took place at 2 single centers and included adults reviewed in 2021 through 2022 from the ATOM1 study (atropine 1% vs placebo; 1999 through 2003) and the ATOM2 study (atropine 0.01% vs 0.1% vs 0.5%; 2006 through 2012). Main Outcome Measures: Change in cycloplegic spherical equivalent (SE) with axial length (AL); incidence of ocular complications. Results: Among the original 400 participants in each original cohort, the study team evaluated 71 of 400 ATOM1 adult participants (17.8% of original cohort; study age, mean [SD] 30.5 [1.2] years; 40.6% female) and 158 of 400 ATOM2 adult participants (39.5% of original cohort; study age, mean [SD], 24.5 [1.5] years; 42.9% female) whose baseline characteristics (SE and AL) were representative of the original cohort. In this study, evaluating ATOM1 participants, the mean (SD) SE and AL were -5.20 (2.46) diopters (D), 25.87 (1.23) mm and -6.00 (1.63) D, 25.90 (1.21) mm in the 1% atropine-treated and placebo groups, respectively (difference of SE, 0.80 D; 95% CI, -0.25 to 1.85 D; P = .13; difference of AL, -0.03 mm; 95% CI, -0.65 to 0.58 mm; P = .92). In ATOM2 participants, the mean (SD) SE and AL was -6.40 (2.21) D; 26.25 (1.34) mm; -6.81 (1.92) D, 26.28 (0.99) mm; and -7.19 (2.87) D, 26.31 (1.31) mm in the 0.01%, 0.1%, and 0.5% atropine groups, respectively. There was no difference in the 20-year incidence of cataract/lens opacities, myopic macular degeneration, or parapapillary atrophy (ß/γ zone) comparing the 1% atropine-treated group vs the placebo group. Conclusions and Relevance: Among approximately one-quarter of the original participants, use of short-term topical atropine eye drops ranging from 0.01% to 1.0% for a duration of 2 to 4 years during childhood was not associated with differences in final refractive errors 10 to 20 years after treatment. There was no increased incidence of treatment or myopia-related ocular complications in the 1% atropine-treated group vs the placebo group. These findings may affect the design of future clinical trials, as further studies are required to investigate the duration and concentration of atropine for childhood myopia control.

Explainable artificial intelligence in emergency medicine: an overview.

Artificial intelligence (AI) and machine learning (ML) have potential to revolutionize emergency medical care by enhancing triage systems, improving diagnostic accuracy, refining prognostication, and optimizing various aspects of clinical care. However, as clinicians often lack AI expertise, they might perceive AI as a "black box," leading to trust issues. To address this, "explainable AI," which teaches AI functionalities to end-users, is important. This review presents the definitions, importance, and role of explainable AI, as well as potential challenges in emergency medicine. First, we introduce the terms explainability, interpretability, and transparency of AI models. These terms sound similar but have different roles in discussion of AI. Second, we indicate that explainable AI is required in clinical settings for reasons of justification, control, improvement, and discovery and provide examples. Third, we describe three major categories of explainability: pre-modeling explainability, interpretable models, and post-modeling explainability and present examples (especially for post-modeling explainability), such as visualization, simplification, text justification, and feature relevance. Last, we show the challenges of implementing AI and ML models in clinical settings and highlight the importance of collaboration between clinicians, developers, and researchers. This paper summarizes the concept of "explainable AI" for emergency medicine clinicians. This review may help clinicians understand explainable AI in emergency contexts.

Handling missing values in healthcare data: A systematic review of deep learning-based imputation techniques.

OBJECTIVE: The proper handling of missing values is critical to delivering reliable estimates and decisions, especially in high-stakes fields such as clinical research. In response to the increasing diversity and complexity of data, many researchers have developed deep learning (DL)-based imputation techniques. We conducted a systematic review to evaluate the use of these techniques, with a particular focus on the types of data, intending to assist healthcare researchers from various disciplines in dealing with missing data. MATERIALS AND METHODS: We searched five databases (MEDLINE, Web of Science, Embase, CINAHL, and Scopus) for articles published prior to February 8, 2023 that described the use of DL-based models for imputation. We examined selected articles from four perspectives: data types, model backbones (i.e., main architectures), imputation strategies, and comparisons with non-DL-based methods. Based on data types, we created an evidence map to illustrate the adoption of DL models. RESULTS: Out of 1822 articles, a total of 111 were included, of which tabular static data (29%, 32/111) and temporal data (40%, 44/111) were the most frequently investigated. Our findings revealed a discernible pattern in the choice of model backbones and data types, for example, the dominance of autoencoder and recurrent neural networks for tabular temporal data. The discrepancy in imputation strategy usage among data types was also observed. The "integrated" imputation strategy, which solves the imputation task simultaneously with downstream tasks, was most popular for tabular temporal data (52%, 23/44) and multi-modal data (56%, 5/9). Moreover, DL-based imputation methods yielded a higher level of imputation accuracy than non-DL methods in most studies. CONCLUSION: The DL-based imputation models are a family of techniques, with diverse network structures. Their designation in healthcare is usually tailored to data types with different characteristics. Although DL-based imputation models may not be superior to conventional approaches across all datasets, it is highly possible for them to achieve satisfactory results for a particular data type or dataset. There are, however, still issues with regard to portability, interpretability, and fairness associated with current DL-based imputation models.

A universal AutoScore framework to develop interpretable scoring systems for predicting common types of clinical outcomes.

The AutoScore framework can automatically generate data-driven clinical scores in various clinical applications. Here, we present a protocol for developing clinical scoring systems for binary, survival, and ordinal outcomes using the open-source AutoScore package. We describe steps for package installation, detailed data processing and checking, and variable ranking. We then explain how to iterate through steps for variable selection, score generation, fine-tuning, and evaluation to generate understandable and explainable scoring systems using data-driven evidence and clinical knowledge. For complete details on the use and execution of this protocol, please refer to Xie et al. (2020),1 Xie et al. (2022)2, Saffari et al. (2022)3 and the online tutorial https://nliulab.github.io/AutoScore/.

Federated and distributed learning applications for electronic health records and structured medical data: a scoping review.

OBJECTIVES: Federated learning (FL) has gained popularity in clinical research in recent years to facilitate privacy-preserving collaboration. Structured data, one of the most prevalent forms of clinical data, has experienced significant growth in volume concurrently, notably with the widespread adoption of electronic health records in clinical practice. This review examines FL applications on structured medical data, identifies contemporary limitations, and discusses potential innovations. MATERIALS AND METHODS: We searched 5 databases, SCOPUS, MEDLINE, Web of Science, Embase, and CINAHL, to identify articles that applied FL to structured medical data and reported results following the PRISMA guidelines. Each selected publication was evaluated from 3 primary perspectives, including data quality, modeling strategies, and FL frameworks. RESULTS: Out of the 1193 papers screened, 34 met the inclusion criteria, with each article consisting of one or more studies that used FL to handle structured clinical/medical data. Of these, 24 utilized data acquired from electronic health records, with clinical predictions and association studies being the most common clinical research tasks that FL was applied to. Only one article exclusively explored the vertical FL setting, while the remaining 33 explored the horizontal FL setting, with only 14 discussing comparisons between single-site (local) and FL (global) analysis. CONCLUSIONS: The existing FL applications on structured medical data lack sufficient evaluations of clinically meaningful benefits, particularly when compared to single-site analyses. Therefore, it is crucial for future FL applications to prioritize clinical motivations and develop designs and methodologies that can effectively support and aid clinical practice and research.

A translational perspective towards clinical AI fairness.

Artificial intelligence (AI) has demonstrated the ability to extract insights from data, but the fairness of such data-driven insights remains a concern in high-stakes fields. Despite extensive developments, issues of AI fairness in clinical contexts have not been adequately addressed. A fair model is normally expected to perform equally across subgroups defined by sensitive variables (e.g., age, gender/sex, race/ethnicity, socio-economic status, etc.). Various fairness measurements have been developed to detect differences between subgroups as evidence of bias, and bias mitigation methods are designed to reduce the differences detected. This perspective of fairness, however, is misaligned with some key considerations in clinical contexts. The set of sensitive variables used in healthcare applications must be carefully examined for relevance and justified by clear clinical motivations. In addition, clinical AI fairness should closely investigate the ethical implications of fairness measurements (e.g., potential conflicts between group- and individual-level fairness) to select suitable and objective metrics. Generally defining AI fairness as "equality" is not necessarily reasonable in clinical settings, as differences may have clinical justifications and do not indicate biases. Instead, "equity" would be an appropriate objective of clinical AI fairness. Moreover, clinical feedback is essential to developing fair and well-performing AI models, and efforts should be made to actively involve clinicians in the process. The adaptation of AI fairness towards healthcare is not self-evident due to misalignments between technical developments and clinical considerations. Multidisciplinary collaboration between AI researchers, clinicians, and ethicists is necessary to bridge the gap and translate AI fairness into real-life benefits.

Shapley variable importance cloud for interpretable machine learning.

Interpretable machine learning has been focusing on explaining final models that optimize performance. The state-of-the-art Shapley additive explanations (SHAP) locally explains the variable impact on individual predictions and has recently been extended to provide global assessments across the dataset. Our work further extends "global" assessments to a set of models that are "good enough" and are practically as relevant as the final model to a prediction task. The resulting Shapley variable importance cloud consists of Shapley-based importance measures from each good model and pools information across models to provide an overall importance measure, with uncertainty explicitly quantified to support formal statistical inference. We developed visualizations to highlight the uncertainty and to illustrate its implications to practical inference. Building on a common theoretical basis, our method seamlessly complements the widely adopted SHAP assessments of a single final model to avoid biased inference, which we demonstrate in two experiments using recidivism prediction data and clinical data.