Multiple Imputation By Chained Equations Research Articles

Estimating Missing Panel Data with Regression and Multivariate Imputation by Chained Equations (MICE)

Missing data may occur in various types of research. Regression and multiple imputation by chained equations (MICE) are two methods that can be used to estimate missing data in panel data types. This study aims to compare the accuracy of the missing panel data estimation using the regression and the MICE methods. The data used in this study are 161 random samples of senior high schools and vocational schools in DKI province for the year 2016-2020. Based on the results of the Chow test, Hausman test, and Lagrange Multiplier test on panel data regression, it shows that the appropriate model for the student-teacher ratio (X5) is random, the percentage of teachers who have an educator certificate (X6) is a fixed model with the specific effect of individual school and time, while the percentage of teachers who hold a bachelor degree (X7) is a fixed model with the specific effect of individual. Based on this model, the estimation of missing data is then carried out. The accuracy of the missing data estimation was carried out by comparing the MAPE, MAE, and RMSE values. The results show that the MICE method is quite good for estimating missing data at X5, quite feasible for estimating X6, and very good for estimating missing data at X7. In general, MICE is more accurate than panel data regression

The hepatocellular carcinoma risk in patients with HBV-related cirrhosis: a competing risk nomogram based on a 4-year retrospective cohort study

ObjectiveThe study aimed to build and validate a competitive risk nomogram to predict the cumulative incidence of hepatocellular carcinoma (HCC) for patients with hepatitis B virus (HBV)-related cirrhosis.MethodsA total of 1401 HBV-related cirrhosis patients were retrospectively enrolled from January 1, 2011 to December 31, 2014. Application of 20 times imputation dealt with missing data using multiple imputation by chained equations (MICE). The patients were randomly divided into a training set (n = 1017) and a validation set (n = 384) at a ratio of 3:1. A prediction study was carried out using a competing risk model, where the event of interest was HCC and the competing events were death and liver transplantation, and subdistribution hazard ratios (sHRs) with 95% CIs were reported. The multivariate competing risk model was constructed and validated.ResultsThere was a negligible difference between the original database and the 20 imputed datasets. At the end of follow-up, the median follow-up time was 69.9 months (interquartile range: 43.8–86.6). There were 31.5% (442/1401) of the patients who developed HCC, with a 5-year cumulative incidence of 22.9 (95%CI, 20.8%–25.2%). The univariate and multivariate competing risk regression and construction of the nomogram were performed in 20 imputed training datasets. Age, sex, antiviral therapy history, hepatitis B e antigen, alcohol drinking history, and alpha-fetoprotein levels were included in the nomogram. The area under receiver operating characteristic curve values at 12, 24, 36, 60, and 96 months were 0.68, 0.69, 0.70, 0.68, and 0.80, and the Brier scores were 0.30, 0.25, 0.23, 0.21, and 0.20 in the validation set. According to the cumulative incidence function, the nomogram effectively screened out high-risk HCC patients from low-risk patients in the presence of competing events (Fine–Gray test p < 0.001).ConclusionThe competitive risk nomogram was allowed to be used for predicting HCC risk in individual patients with liver cirrhosis, taking into account both the association between risk factors and HCC and the modifying effect of competition events on this association.

Quality of life and quality-adjusted life years after stroke in Sierra Leone.

Stroke is a leading cause of mortality and negatively affects health-related quality of life (HRQoL). HRQoL after stroke is understudied in Africa and there are no reports of quality-adjusted life years after stroke (QALYs) in African countries. We determined the impact of stroke on HRQoL after stroke in Sierra Leone. We calculated QALYs at 1 year post-stroke and determined sociodemographic and clinical variables associated with HRQoL and QALYs in this population. A prospective stroke register was established at the two-principal adult tertiary government hospitals in Freetown, Sierra Leone. Participants were followed up at 7, 90 days, and 1 year post-stroke to capture all-cause mortality and EQ-5D-3L data. QALYs were calculated at the patient level using EQ-5D-3L utility values and survival data from the register, following the area under the curve method. Utilities were based on the UK and Zimbabwe (as a sensitivity analysis) EQ-5D value sets, as there is no Sierra Leonean or West African value set. Explanatory models were developed based on previous literature to assess variables associated with HRQoL and QALYs at 1 year after stroke. To address missing values, Multiple Imputation by Chained Equations (MICE), with linear and logistic regression models for continuous and binary variables, respectively, were used. EQ-5D-3L data were available for 373/460 (81.1%), 360/367 (98.1%), and 299/308 (97.1%) participants at 7, 90 days, and 1 year after stroke. For stroke survivors, median EQ-5D-3L utility increased from 0.20 (95% CI: -0.16 to 0.59) at 7 days post-stroke to 0.76 (0.47 to 1.0) at 90 days and remained stable at 1 year 0.76 (0.49 to 1.0). Mean QALYs at 1 year after stroke were 0.28 (SD: 0.35) and closely associated with stroke severity. Older age, lower educational attainment, patients with subarachnoid hemorrhage and undetermined stroke types all had lower QALYs and lower HRQoL, while being the primary breadwinner was associated with higher HRQoL. Sensitivity analysis with the Zimbabwe value set did not significantly change regression results but did influence the absolute values with Zimbabwe utility values being higher, with fewer utility values less than 0. We generated QALYs after stroke for the first time in an African country. QALYs were significantly lower than studies from outside Africa, partially explained by the high mortality rate in our cohort. Further research is needed to develop appropriate value sets for West African countries and to examine QALYs lost due to stroke over longer time periods. The Stroke in Sierra Leone anonymized dataset is available on request to researchers, see data access section.

Improving Regression Analysis with Imputation in a Longitudinal Study of Alzheimer's Disease.

Missing data is prevalent in the Alzheimer's Disease Neuroimaging Initiative (ADNI). It is common to deal with missingness by removing subjects with missing entries prior to statistical analysis; however, this can lead to significant efficiency loss and sometimes bias. It has yet to be demonstrated that the imputation approach to handling this issue can be valuable in some longitudinal regression settings. The purpose of this study is to demonstrate the importance of imputation and how imputation is correctly done in ADNI by analyzing longitudinal Alzheimer's Disease Assessment Scale -Cognitive Subscale 13 (ADAS-Cog 13) scores and their association with baseline patient characteristics. We studied 1,063 subjects in ADNI with mild cognitive impairment. Longitudinal ADAS-Cog 13 scores were modeled with a linear mixed-effects model with baseline clinical and demographic characteristics as predictors. The model estimates obtained without imputation were compared with those obtained after imputation with Multiple Imputation by Chained Equations (MICE). We justify application of MICE by investigating the missing data mechanism and model assumptions. We also assess robustness of the results to the choice of imputation method. The fixed-effects estimates of the linear mixed-effects model after imputation with MICE yield valid, tighter confidence intervals, thus improving the efficiency of the analysis when compared to the analysis done without imputation. Our study demonstrates the importance of accounting for missing data in ADNI. When deciding to perform imputation, care should be taken in choosing the approach, as an invalid one can compromise the statistical analyses.

The future burden of type 2 diabetes in Belgium: a microsimulation model

ObjectiveTo forecast the annual burden of type 2 diabetes and related socio-demographic disparities in Belgium until 2030.MethodsThis study utilized a discrete-event transition microsimulation model. A synthetic population was created using 2018 national register data of the Belgian population aged 0–80 years, along with the national representative prevalence of diabetes risk factors obtained from the latest (2018) Belgian Health Interview and Examination Surveys using Multiple Imputation by Chained Equations (MICE) as inputs to the Simulation of Synthetic Complex Data (simPop) model. Mortality information was obtained from the Belgian vital statistics and used to calculate annual death probabilities. From 2018 to 2030, synthetic individuals transitioned annually from health to death, with or without developing type 2 diabetes, as predicted by the Finnish Diabetes Risk Score, and risk factors were updated via strata-specific transition probabilities.ResultsA total of 6722 [95% UI 3421, 11,583] new cases of type 2 diabetes per 100,000 inhabitants are expected between 2018 and 2030 in Belgium, representing a 32.8% and 19.3% increase in T2D prevalence rate and DALYs rate, respectively. While T2D burden remained highest for lower-education subgroups across all three Belgian regions, the highest increases in incidence and prevalence rates by 2030 are observed for women in general, and particularly among Flemish women reporting higher-education levels with a 114.5% and 44.6% increase in prevalence and DALYs rates, respectively. Existing age- and education-related inequalities will remain apparent in 2030 across all three regions.ConclusionsThe projected increase in the burden of T2D in Belgium highlights the urgent need for primary and secondary preventive strategies. While emphasis should be placed on the lower-education groups, it is also crucial to reinforce strategies for people of higher socioeconomic status as the burden of T2D is expected to increase significantly in this population segment.

On the search for efficient face recognition algorithm subject to multiple environmental constraints

From literature, majority of face recognition modules suffer performance challenges when presented with test images acquired under multiple constrained environments (occlusion and varying expressions). The performance of these models further deteriorates as the degree of degradation of the test images increases (relatively higher occlusion level). Deep learning-based face recognition models have attracted much attention in the research community as they are purported to outperform the classical PCA-based methods. Unfortunately their application to real-life problems is limited because of their intensive computational complexity and relatively longer run-times. This study proposes an enhancement of some PCA-based methods (with relatively lower computational complexity and run-time) to overcome the challenges posed to the recognition module in the presence of multiple constraints. The study compared the performance of enhanced classical PCA-based method (HE-GC-DWT-PCA/SVD) to FaceNet algorithm (deep learning method) using expression variant face images artificially occluded at 30% and 40%. The study leveraged on two statistical imputation methods of MissForest and Multiple Imputation by Chained Equations (MICE) for occlusion recovery. From the numerical evaluation results, although the two models achieved the same recognition rate (85.19%) at 30% level of occlusion, the enhanced PCA-based algorithm (HE-GC-DWT-PCA/SVD) outperformed the FaceNet model at 40% occlusion rate, with a recognition rate of 83.33%. Although both Missforest and MICE performed creditably well as de-occlusion mechanisms at higher levels of occlusion, MissForest outperforms the MICE imputation mechanism. MissForest imputation mechanism and the proposed HE-GC-DWT-PCA/SVD algorithm are recommended for occlusion recovery and recognition of multiple constrained test images respectively.

The Vitamin D Receptor as a Prognostic Marker in Breast Cancer-A Cohort Study.

Previous research has indicated an association between the presence of the vitamin D receptor (VDR) in breast cancer tissue and a favorable prognosis. This study aimed to further evaluate the prognostic potential of VDR located in the nuclear membrane or nucleus (liganded). The VDR protein levels were analyzed using immunohistochemistry in tumor samples from 878 breast cancer patients from Lund, Sweden, included in the Breast Cancer and Blood Study (BCBlood) from October 2002 to June 2012. The follow-up for breast cancer events and overall survival was recorded until 30 June 2019. Univariable and multivariable survival analyses were conducted, both with complete case data and with missing data imputed using multiple imputation by chained equations (MICE). Tumor-specific positive nuclear membrane VDR(num) staining was associated with favorable tumor characteristics and a longer breast cancer free interval (BCFI; HR: 0.64; 95% CI: 0.44-0.95) and overall survival (OS; HR: 0.52; 95% CI: 0.34-0.78). Further analyses indicated that VDRnum status also was predictive of overall survival when investigated in relation to ER status. There were significant interactions between VDR and invasive tumor size (Pinteraction = 0.047), as well as mode of detection (Pinteraction = 0.049). VDRnum was associated with a longer BCFI in patients with larger tumors (HR: 0.36; 95% CI: 0.14-0.93) or clinically detected tumors (HR: 0.28; 95% CI: 0.09-0.83), while no association was found for smaller tumors and screening-detected tumors. Further studies are suggested to confirm our results and to evaluate whether VDR should and could be used as a prognostic and targetable marker in breast cancer diagnostics.

Handling missing values in mixed panel data: a comparison of different techniques

Purpose- The purpose of this study is to compare the success of alternative data imputation techniques with missing data. The study distinguishes itself from the rest of the literature by proposing an appropriate technique for mixed data on financial performance and environmental, social and governance (ESG) metrics of companies. In addition to simple imputation techniques, we also use machine learning techniques that allow working with more complex data. Methodology- We first employ ad-hoc methods such as mean, median, mode, constant, most frequent and regression imputation. In what follows, we handle multivariate imputation techniques such as multiple imputation by chained equations (MICE). Finally, we run imputation methods with machine learning (ML) classification such as K-nearest Neighbor (KNN), Ridge and Random Forest. To consider the assumptions of missing data, we first check the normality of the variables with Kolmogorov-Smirnov test and employ Rubin’s classification technique that defines the relationship among variables with the probability of missing data. The success of imputation techniques applied to missing data changes when the missing data are classified with Rubin’s technique according to randomness. Consequently, we apply listwise deletion at various levels and alternative data imputation techniques. We then compare their performances. The raw data contain parametric as well as categorical variables (binary and others). Among these are time-series (yearly) financial series such as sales and total assets obtained from financial statements, ESG scores as well as float ratios for firms from several countries and industries. Imputation is done randomly on a sample varying from 5% to 30% of the dataset and results are compared to true data based on accuracy or other measures such as root mean square errors (RMSE) or mean absolute percentage error (MAPE). Several robustness checks have been performed to supplement the analysis. Findings- Results show that ML methods such as KNN have a superior performance than others. Moreover, when multidimensional nature of the data is taken into account, the prediction performance improves. Hence, an optimality can be reached based on parameters. Conclusion- Based upon the analysis, we conclude that the selected imputation technique and how it is employed matter to attain a higher accuracy and a better prediction of the missing values on selected mixed panel data in finance. Keywords: Imputation techniques, Panel data, Machine learning, Financial performance, ESG JEL Codes: C55, C81, M14, Q51

Abstract 14627: Machine Learning Based Models for Imputing Heart Failure on First Medical Contact From the Get With The Guidelines®- Coronary Artery Disease (GWTG-CAD) Registry

Introduction: The American Heart Association Get With the Guidelines - Coronary Artery Disease (GWTG-CAD) Registry is a quality improvement program with over 2600 participating U.S. hospitals. Among the collected data elements is the yes-no variable “Heart failure documented on First Medical Contact” (HF-FMC), an important clinical predictor of in-hospital mortality. Before 2018, the missing rate of HF-FMC exceeded 60% in some years, significantly limiting its utility as a predictor of in-hospitality mortality. We compared machine learning (ML)-based with traditional statistical methods to accurately impute missing HF-FMC values. Methods: We utilized admission records from the GWTG CAD Registry data from 2018 to 2021, including both STEMI and NSTEMI patients. We removed missing HF-FMC, using complete data as ground truth. The data was split into training and test datasets in a 3:1 ratio. Imputation was framed as a prediction task. For evaluation, we masked HF-FMC in the test set and compared the imputed values against the ground truth. For traditional methods, mode imputation, K-nearest neighbor imputer (KNN), and Multiple Imputation by Chained Equations (MICE) were applied. For ML models, we applied logistic regression, random forest and XGBoost. Shapley Additive exPlanations (SHAP) values were utilized to identify key features contributing to the ML models. Results: The study included 223,074 individuals (12.2% HF-FMC = “yes”), with mean (SD) age 64.9 (13.4) years (33.8% Female; and 77.3% White and 13.8% Black). Machine Learning models achieved higher AUROC scores compared to traditional models in imputing HF-FMC (Table 1). Top features for predicting HF-FMC are shown (Figure 1). Conclusions: Machine learning models outperformed traditional methods in imputing missing HF-FMC, a critical predictor of in-hospital mortality. SHAP values can identify key features in imputing clinically important variables with high degree of missingness.

MICE Implementation to Handle Missing Values in Rain Potential Prediction Using Support Vector Machine Algorithm

Support Vector Machine (SVM) is a machine learning algorithm used for classification. SVM has several advantages such as the ability to handle high-dimensional data, effective in handling nonlinear data through kernel functions, and resistance to overfitting through soft margins. However, SVM has weaknesses, especially when handling missing values in data. The use of SVM must consider the missing values strategy chosen. Missing values in data mining is a serious problem for researchers because it causes many problems such as loss of efficiency, complications in data handling and analysis, and the occurrence of bias due to differences between missing data and complete data. To overcome the above problems, this research focuses on understanding the characteristics of missing values and handling them using the Multiple Imputation by Chained Equations (MICE) technique. In this study, we utilized secondary data experiments that contain missing values from the Meteorological, Climatological, and Geophysical Agency (called BMKG) related to predictions of potential rain, especially in DKI Jakarta. Identification of types or patterns of missing values, exploration of the relationship between missing values and other variables, incorporation of the MICE method to handle missing values, and the Support Vector Machine Algorithm for classification will be carried out to produce a more reliable and accurate prediction model for rain potential. It shows that the imputation method with the MICE gives better results than other techniques (such as Complete Case Analysis, Imputation Method Mean, Median, Mode, and K-Nearest neighbor), namely an accuracy of 89% testing data when applying the Support Vector Machine algorithm for classification.

Comparing supervised and semi-supervised machine learning approaches in NTCP modeling to predict complications in head and neck cancer patients

Background and purposeHead and neck cancer (HNC) patients treated with radiotherapy often suffer from radiation-induced toxicities. Normal Tissue Complication Probability (NTCP) modeling can be used to determine the probability to develop these toxicities based on patient, tumor, treatment and dose characteristics. Since the currently used NTCP models are developed using supervised methods that discard unlabeled patient data, we assessed whether the addition of unlabeled patient data by using semi-supervised modeling would gain predictive performance. Materials and methodsThe semi-supervised method of self-training was compared to supervised regression methods with and without prior multiple imputation by chained equation (MICE). The models were developed for the most common toxicity outcomes in HNC patients, xerostomia (dry mouth) and dysphagia (difficulty swallowing), measured at six months after treatment, in a development cohort of 750 HNC patients. The models were externally validated in a validation cohort of 395 HNC patients. Model performance was assessed by discrimination and calibration. ResultsMICE and self-training did not improve performance in terms of discrimination or calibration at external validation compared to current regression models. In addition, the relative performance of the different models did not change upon a decrease in the amount of (labeled) data available for model development. Models using ridge regression outperformed the logistic models for the dysphagia outcome. ConclusionSince there was no apparent gain in the addition of unlabeled patient data by using the semi-supervised method of self-training or MICE, the supervised regression models would still be preferred in current NTCP modeling for HNC patients.

Data augmentation for bias correction in mapping PM2.5 based on satellite retrievals and ground observations

As most air quality monitoring sites are in urban areas worldwide, machine learning models may produce substantial estimation bias in rural areas when deriving spatiotemporal distributions of air pollutants. The bias stems from the issue of dataset shift, as the density distributions of predictor variables differ greatly between urban and rural areas. We propose a data-augmentation approach based on the multiple imputation by chained equations (MICE-DA) to remedy the dataset shift problem. Compared with the benchmark models, MICE-DA exhibits superior predictive performance in deriving the spatiotemporal distributions of hourly PM2.5 in the megacity (Chengdu) at the foot of the Tibetan Plateau, especially for correcting the estimation bias, with the mean bias decreasing from –3.4 µg/m3 to –1.6 µg/m3. As a complement to the holdout validation, the semi-variance results show that MICE-DA decently preserves the spatial autocorrelation pattern of PM2.5 over the study area. The essence of MICE-DA is strengthening the correlation between PM2.5 and aerosol optical depth (AOD) during the data augmentation. Consequently, the importance of AOD is largely enhanced for predicting PM2.5, and the summed relative importance value of the two satellite-retrieved AOD variables increases from 5.5% to 18.4%. This study resolved the puzzle that AOD exhibited relatively lower importance in local or regional studies. The results of this study can advance the utilization of satellite remote sensing in modeling air quality while drawing more attention to the common dataset shift problem in data-driven environmental research.

Neural Network Approach to Investigating the Importance of Test Items for Predicting Physical Activity in Chronic Obstructive Pulmonary Disease.

Contracting COPD reduces a patient's physical activity and restricts everyday activities (physical activity disorder). However, the fundamental cause of physical activity disorder has not been found. In addition, costly and specialized equipment is required to accurately examine the disorder; hence, it is not regularly assessed in normal clinical practice. In this study, we constructed a machine learning model to predict physical activity using test items collected during the normal care of COPD patients. In detail, we first applied three types of data preprocessing methods (zero-padding, multiple imputation by chained equations (MICE), and k-nearest neighbor (kNN)) to complement missing values in the dataset. Then, we constructed several types of neural networks to predict physical activity. Finally, permutation importance was calculated to identify the importance of the test items for prediction. Multifactorial analysis using machine learning, including blood, lung function, walking, and chest imaging tests, was the unique point of this research. From the experimental results, it was found that the missing value processing using MICE contributed to the best prediction accuracy (73.00%) compared to that using zero-padding (68.44%) or kNN (71.52%), and showed better accuracy than XGBoost (66.12%) with a significant difference (p < 0.05). For patients with severe physical activity reduction (total exercise < 1.5), a high sensitivity (89.36%) was obtained. The permutation importance showed that "sex, the number of cigarettes, age, and the whole body phase angle (nutritional status)" were the most important items for this prediction. Furthermore, we found that a smaller number of test items could be used in ordinary clinical practice for the screening of physical activity disorder.

Responsible Imputation of Missing Speech Perception Testing Data & Analysis of 4,739 Observations and Predictors of Performance.

To address outcome heterogeneity in cochlear implant (CI) research, we built imputation models using multiple imputation by chained equations (MICEs) and K-nearest neighbors (KNNs) to convert between four common open-set testing scenarios: Consonant-Nucleus-Consonant word (CNCw), Arizona Biomedical (AzBio) in quiet, AzBio +5, and AzBio +10. We then analyzed raw and imputed data sets to evaluate factors affecting CI outcome variability. Retrospective cohort study of a national CI database (HERMES) and a nonoverlapping single-institution CI database. Multi-institutional (32 CI centers). Adult CI recipients (n = 4,046 patients). Mean absolute error (MAE) between imputed and observed speech perception scores. Imputation models of preoperative speech perception measures demonstrate a MAE of less than 10% for feature triplets of CNCw/AzBio in quiet/AzBio +10 (MICE: MAE, 9.52%; 95% confidence interval [CI], 9.40-9.64; KNN: MAE, 8.93%; 95% CI, 8.83-9.03) and AzBio in quiet/AzBio +5/AzBio +10 (MICE: MAE, 8.85%; 95% CI, 8.68-9.02; KNN: MAE, 8.95%; 95% CI, 8.74-9.16) with one feature missing. Postoperative imputation can be safely performed with up to four of six features missing in a set of CNCw and AzBio in quiet at 3, 6, and 12 months postcochlear implantation using MICE (MAE, 9.69%; 95% CI, 9.63-9.76). For multivariable analysis of CI performance prediction, imputation increased sample size by 72%, from 2,756 to 4,739, with marginal change in adjusted R2 (0.13 raw, 0.14 imputed). Missing data across certain sets of common speech perception tests may be safely imputed, enabling multivariate analysis of one of the largest CI outcomes data sets to date.

Abstract 4174: Incorporating continuous mammographic density into the BOADICEA breast cancer risk prediction model

Abstract BOADICEA, implemented in the CanRisk tool (www.canrisk.org) can be used to calculate future breast cancer risk using data on cancer family history, genetics (rare high- or moderate-risk pathogenic variants; polygenic scores (PGS)), questionnaire-based risk factors and mammographic density (MD) measured using the BI-RADS classification. The BI-RADS categorization requires manual reading which is not feasible at population level. Moreover, BI-RADS captures MD in 4 categories, whereas continuous measures are stronger risk predictors. Here, BOADICEA was extended to incorporate continuous breast density measured by the fully automated Volpara and STRATUS software, working on raw and processed images, respectively. We used data from the KARMA prospective screening cohort (60,276 participants; 1,167 incident breast cancers). The associations between Volpara or STRATUS density measurements and incident breast cancer risk were estimated in a randomly selected training subset consisting of two-thirds of the full dataset. For this, we calculated density residuals after regressing on participant’s age at mammography. Hazard ratios (HR) for the normalized residuals were calculated using a Cox proportional hazards model, adjusting for family history, BMI and all risk factors included in BOADICEA. Multiple Imputation by Chained Equations (MICE) was used to impute missing data. The remaining one-third of the KARMA cohort was used to assess the performance of BOADICEA in predicting 5-year risks, after including the estimated Volpara and STRATUS associations.Under the best fitting multivariable model including PGS, the HRs per SD of residual STRATUS density were estimated to be 1.67 (95%CI: 1.44-1.93) and 1.38 (95%CI: 1.26-1.51) for pre- and post-menopausal women, respectively. The HR estimates per SD of residual Volpara density were 1.38 (95%CI: 1.19-1.59) and 1.39 (95%CI: 1.27-1.52) for pre- and post-menopausal women, respectively. After incorporating these associations in BOADICEA, there was a marked improvement in the model discriminatory ability in the test dataset, compared to using the BI-RADS classification. The largest increase in AUC was observed for STRATUS residual density, with a 3% increase compared to using BI-RADS vs a 1.6% increase in AUC when using Volpara residual density. These increases were consistent across BOADICEA models considering different combinations of risk factors, including PGS.Including continuous breast density in BOADICEA can lead to improved breast cancer risk stratification and could allow for automated measures of MD to be more readily deployed in breast cancer risk prediction. Additional prospective studies are required to further validate the findings. The models will be implemented in the CanRisk tool. Citation Format: Lorenzo Ficorella, Mikael Eriksson, Kamila Czene, Goska Leslie, Xin Yang, Tim J. Carver, Douglas F. Easton, Per F. Hall, Antonis C. Antoniou. Incorporating continuous mammographic density into the BOADICEA breast cancer risk prediction model. [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 4174.

Improving Cardiovascular Disease Prediction by Integrating Imputation, Imbalance Resampling, and Feature Selection Techniques into Machine Learning Model

Cardiovascular disease (CVD) is the leading cause of death worldwide. Primary prevention is by early prediction of the disease onset. Using laboratory data from the National Health and Nutrition Examination Survey (NHANES) in 2017-2020 timeframe (N= 7.974), we tested the ability of machine learning (ML) algorithms to classify individuals at risk. The ML models were evaluated based on their classification performances after comparing four imputation, three imbalance resampling, and three feature selection techniques.Due to its popularity, we utilized decision tree (DT) as the baseline. Integration of multiple imputation by chained equation (MICE) and synthetic minority oversampling with Tomek link down-sampling (SMOTETomek) into the model improved the area under the curve-receiver operating characteristics (AUC-ROC) from 57% to 83%. Applying simultaneous perturbation feature selection and ranking (spFSR) reduced the feature predictors from 144 to 30 features and the computational time by 22%. The best techniques were applied to six ML models, resulting in Xtreme gradient boosting (XGBoost) achieving the highest accuracy of 93% and AUC-ROC of 89%.The accuracy of our ML model in predicting CVD outperforms those from previous studies. We also highlight the important causes of CVD, which might be investigated further for potential effects on electronic health records.

Imputation of missing values for cochlear implant candidate audiometric data and potential applications.

Assess the real-world performance of popular imputation algorithms on cochlear implant (CI) candidate audiometric data. 7,451 audiograms from patients undergoing CI candidacy evaluation were pooled from 32 institutions with complete case analysis yielding 1,304 audiograms. Imputation model performance was assessed with nested cross-validation on randomly generated sparse datasets with various amounts of missing data, distributions of sparsity, and dataset sizes. A threshold for safe imputation was defined as root mean square error (RMSE) <10dB. Models included univariate imputation, interpolation, multiple imputation by chained equations (MICE), k-nearest neighbors, gradient boosted trees, and neural networks. Greater quantities of missing data were associated with worse performance. Sparsity in audiometric data is not uniformly distributed, as inter-octave frequencies are less commonly tested. With 3-8 missing features per instance, a real-world sparsity distribution was associated with significantly better performance compared to other sparsity distributions (Δ RMSE 0.3 dB- 5.8 dB, non-overlapping 99% confidence intervals). With a real-world sparsity distribution, models were able to safely impute up to 6 missing datapoints in an 11-frequency audiogram. MICE consistently outperformed other models across all metrics and sparsity distributions (p < 0.01, Wilcoxon rank sum test). With sparsity capped at 6 missing features per audiogram but otherwise equivalent to the raw dataset, MICE imputed with RMSE of 7.83 dB [95% CI 7.81-7.86]. Imputing up to 6 missing features captures 99.3% of the audiograms in our dataset, allowing for a 5.7-fold increase in dataset size (1,304 to 7,399 audiograms) as compared with complete case analysis. Precision medicine will inevitably play an integral role in the future of hearing healthcare. These methods are data dependent, and rigorously validated imputation models are a key tool for maximizing datasets. Using the largest CI audiogram dataset to-date, we demonstrate that in a real-world scenario MICE can safely impute missing data for the vast majority (>99%) of audiograms with RMSE well below a clinically significant threshold of 10dB. Evaluation across a range of dataset sizes and sparsity distributions suggests a high degree of generalizability to future applications.

A comparison of imputation methods for categorical data

ObjectivesMissing data is commonplace in clinical databases, which are being increasingly used for research. Without giving any regard to missing data, results from analysis may become biased and unrepresentative. Clinical databases contain mainly categorical variables. This study aims to assess the methods used for imputation in categorical variables. Materials and methodsWe utilized data extracted from paper-based maternal health records from Kawempe National Referral Hospital, Uganda. We compared the following imputation methods for categorical data in an empirical analysis: Mode, K-Nearest Neighbors (KNN), Random Forest (RF), Sequential Hot-Deck (SHD), and Multiple Imputation by Chained Equations (MICE). The five imputation methods were first compared by accuracy of predicting the missing values. Next, the imputation methods were compared by predictive accuracy of the outcome variable in four classifiers. The consistency of performance of imputation methods across different levels of missing data (5%–50 %) was assessed by Kendall's W test. ResultsKNN imputation had the highest precision score at levels (5%–50 %) of MCAR missing data. At lower proportions of missing data (5 %, 10 %, 15 %, 20 %), RF imputation had the second-highest precision score. SHD imputation had the worst precision at all levels of missing data. In the prediction of the outcome, the methods performed differently at all proportions of missing data in the four classifiers. Even though KNN imputation was the best method in predicting the missing values, it did not consistently enhance the predictive accuracy of the classifiers at all levels of missing data. Our findings show that a high precision score of an imputation method does not translate into higher predictive accuracy in classifiers. ConclusionsKNN imputation is the best method in predicting missing values in categorical variables. There is no universal best imputation method that yields the highest predictive accuracy at all proportions of missing data.

Treatment of missing data in Bayesian network structure learning: an application to linked biomedical and social survey data

BackgroundAvailability of linked biomedical and social science data has risen dramatically in past decades, facilitating holistic and systems-based analyses. Among these, Bayesian networks have great potential to tackle complex interdisciplinary problems, because they can easily model inter-relations between variables. They work by encoding conditional independence relationships discovered via advanced inference algorithms. One challenge is dealing with missing data, ubiquitous in survey or biomedical datasets. Missing data is rarely addressed in an advanced way in Bayesian networks; the most common approach is to discard all samples containing missing measurements. This can lead to biased estimates. Here, we examine how Bayesian network structure learning can incorporate missing data.MethodsWe use a simulation approach to compare a commonly used method in frequentist statistics, multiple imputation by chained equations (MICE), with one specific for Bayesian network learning, structural expectation-maximization (SEM). We simulate multiple incomplete categorical (discrete) data sets with different missingness mechanisms, variable numbers, data amount, and missingness proportions. We evaluate performance of MICE and SEM in capturing network structure. We then apply SEM combined with community analysis to a real-world dataset of linked biomedical and social data to investigate associations between socio-demographic factors and multiple chronic conditions in the US elderly population.ResultsWe find that applying either method (MICE or SEM) provides better structure recovery than doing nothing, and SEM in general outperforms MICE. This finding is robust across missingness mechanisms, variable numbers, data amount and missingness proportions. We also find that imputed data from SEM is more accurate than from MICE. Our real-world application recovers known inter-relationships among socio-demographic factors and common multimorbidities. This network analysis also highlights potential areas of investigation, such as links between cancer and cognitive impairment and disconnect between self-assessed memory decline and standard cognitive impairment measurement.ConclusionOur simulation results suggest taking advantage of the additional information provided by network structure during SEM improves the performance of Bayesian networks; this might be especially useful for social science and other interdisciplinary analyses. Our case study show that comorbidities of different diseases interact with each other and are closely associated with socio-demographic factors.

Handling missing data in well-log curves with a gated graph neural network

Well logging is a common method that is used to obtain the rock properties of a formation. It is relatively frequent, however, that log information is incomplete due to cost limitations or borehole problems. Existing models predict missing well logs from a fixed combination of other available well logs. However, the missing well logs vary from well to well. We have proposed using a gated graph neural network (GNN) to handle the missing values in well-log curves. It takes sequential data, predicting each missing measurement in the data not only using other available variables measured at the same depth but also available measurements of neighboring observations. Meanwhile, the missing well logs and available well logs could be any possible combinations as long as they are mutually exclusive. This approach has two advantages: (1) the gated GNN does not need to build a specific model for each missing measurement or from every possible combination of available measurements and (2) it can be integrated into the training process of the following predictive model to perform classification tasks. We evaluate the gated GNN model along with two other models: the GRAPE model and the multiple imputation by chained equations (MICE)-gated recurrent unit (GRU) model, on a data set from the North Sea to perform a missing feature imputation task and a lithofacies identification task. The GRAPE model also is a graph-based model, and it predicts values for each missing measurement from available variables measured at the same depth. The MICE-GRU model is a combination of the MICE algorithm and GRU, which handles the feature imputation procedure and the lithofacies identification procedure separately. Our experiments find that the gated GNN model outperforms the MICE algorithm and the GRAPE model on the missing feature imputation task. For the lithofacies identification task, the gated GNN model also provides comparable results to the MICE-GRU model, and they both outperform the GRAPE model.