Machine learning approaches to identify systemic lupus erythematosus in anti-nuclear antibody-positive patients using genomic data and electronic health records.

BACKGROUND: Although the 2019 EULAR/ACR classification criteria for systemic lupus erythematosus (SLE) has required at least a positive anti-nuclear antibody (ANA) titer (≥ 1:80), it remains challenging for clinicians to identify patients with SLE. This study aimed to develop a machine learning (ML) approach to assist in the detection of SLE patients using genomic data and electronic health records. METHODS: Participants with a positive ANA (≥ 1:80) were enrolled from the Taiwan Precision Medicine Initiative cohort. The Taiwan Biobank version 2 array was used to detect single nucleotide polymorphism (SNP) data. Six ML models, Logistic Regression, Random Forest (RF), Support Vector Machine, Light Gradient Boosting Machine, Gradient Tree Boosting, and Extreme Gradient Boosting (XGB), were used to identify SLE patients. The importance of the clinical and genetic features was determined by Shapley Additive Explanation (SHAP) values. A logistic regression model was applied to identify genetic variations associated with SLE in the subset of patients with an ANA equal to or exceeding 1:640. RESULTS: A total of 946 SLE and 1,892 non-SLE controls were included in this analysis. Among the six ML models, RF and XGB demonstrated superior performance in the differentiation of SLE from non-SLE. The leading features in the SHAP diagram were anti-double strand DNA antibodies, ANA titers, AC4 ANA pattern, polygenic risk scores, complement levels, and SNPs. Additionally, in the subgroup with a high ANA titer (≥ 1:640), six SNPs positively associated with SLE and five SNPs negatively correlated with SLE were discovered. CONCLUSIONS: ML approaches offer the potential to assist in diagnosing SLE and uncovering novel SNPs in a group of patients with autoimmunity.

Machine learning approaches for the genomic prediction of rheumatoid arthritis and systemic lupus erythematosus.

BACKGROUND: Rheumatoid arthritis (RA) and systemic lupus erythematous (SLE) are autoimmune rheumatic diseases that share a complex genetic background and common clinical features. This study's purpose was to construct machine learning (ML) models for the genomic prediction of RA and SLE. METHODS: A total of 2,094 patients with RA and 2,190 patients with SLE were enrolled from the Taichung Veterans General Hospital cohort of the Taiwan Precision Medicine Initiative. Genome-wide single nucleotide polymorphism (SNP) data were obtained using Taiwan Biobank version 2 array. The ML methods used were logistic regression (LR), random forest (RF), support vector machine (SVM), gradient tree boosting (GTB), and extreme gradient boosting (XGB). SHapley Additive exPlanation (SHAP) values were calculated to clarify the contribution of each SNPs. Human leukocyte antigen (HLA) imputation was performed using the HLA Genotype Imputation with Attribute Bagging package. RESULTS: Compared with LR (area under the curve [AUC] = 0.8247), the RF approach (AUC = 0.9844), SVM (AUC = 0.9828), GTB (AUC = 0.9932), and XGB (AUC = 0.9919) exhibited significantly better prediction performance. The top 20 genes by feature importance and SHAP values included HLA class II alleles. We found that imputed HLA-DQA1*05:01, DQB1*0201 and DRB1*0301 were associated with SLE; HLA-DQA1*03:03, DQB1*0401, DRB1*0405 were more frequently observed in patients with RA. CONCLUSIONS: We established ML methods for genomic prediction of RA and SLE. Genetic variations at HLA-DQA1, HLA-DQB1, and HLA-DRB1 were crucial for differentiating RA from SLE. Future studies are required to verify our results and explore their mechanistic explanation.

Application of Supervised Machine Learning to Recognize Competent Level and Mixed Antinuclear Antibody Patterns Based on ICAP International Consensus.

BACKGROUND: Antinuclear antibody pattern recognition is vital for autoimmune disease diagnosis but labor-intensive for manual interpretation. To develop an automated pattern recognition system, we established machine learning models based on the International Consensus on Antinuclear Antibody Patterns (ICAP) at a competent level, mixed patterns recognition, and evaluated their consistency with human reading. METHODS: 51,694 human epithelial cells (HEp-2) cell images with patterns assigned by experienced medical technologists collected in a medical center were used to train six machine learning algorithms and were compared by their performance. Next, we choose the best performing model to test the consistency with five experienced readers and two beginners. RESULTS: The mean F1 score in each classification of the best performing model was 0.86 evaluated by Testing Data 1. For the inter-observer agreement test on Testing Data 2, the average agreement was 0.849 (κ) among five experienced readers, 0.844 between the best performing model and experienced readers, 0.528 between experienced readers and beginners. The results indicate that the proposed model outperformed beginners and achieved an excellent agreement with experienced readers. CONCLUSIONS: This study demonstrated that the developed model could reach an excellent agreement with experienced human readers using machine learning methods.

First-4-week erythrocyte sedimentation rate variability predicts erythrocyte sedimentation rate trajectories and clinical course among patients with pyogenic vertebral osteomyelitis.

BACKGROUND: The trajectory pattern of erythrocyte sedimentation rate (ESR) in patients with pyogenic vertebral osteomyelitis (PVO) and its clinical significance is unclear. We further evaluated whether the first-4-week ESR variability can predict the trajectory pattern, treatment duration and recurrence of PVO. METHODS: The longitudinal ESR patterns of adults with PVO within the first 6 months were characterized through group-based trajectory modeling (GBTM). The ESR variability within the first 4 weeks was defined using the absolute difference (AD), coefficient of variation, percent change, and slope change. The first-4-week ESR variabilities were analyzed using ordinal logistic regression to predict the 6-month ESR trajectory and using logistic regression to predict treatment duration and recurrence likelihood. The discrimination and calibration of the prediction models were evaluated. RESULTS: Three ESR trajectory patterns were identified though GBTM among patients with PVO: Group 1, initial moderate high ESR with fast response; Group 2, initial high ESR with fast response; Group 3, initial high ESR with slow response. Group 3 patients (initial high ESR with slow response) were older, received longer antibiotic treatment, and had more comorbidities and higher recurrence rates than patients in the other two groups. The initial ESR value and ESR - AD could predict the 6-month ESR trajectory. By incorporating the first-4-week ESR variabilities and the clinical features of patients, our models exhibited moderate discrimination performance to predict prolonged treatment (≥12 weeks; C statistic, 0.75; 95% confidence interval [CI], 0.70 to 0.81) and recurrence (C statistic, 0.69; 95% CI, 0.61 to 0.78). CONCLUSIONS: The initial ESR value and first-4-week ESR variability are useful markers to predict the treatment duration and recurrence of PVO. Future studies should validate our findings in other populations.

Risk prediction for 30-day mortality among patients with Clostridium difficile infections: a retrospective cohort study.

Background: Current guidelines have unsatisfied performance in predicting severe outcomes after Clostridium difficile infection (CDI). Our objectives were to develop a risk prediction model for 30-day mortality and to examine its performance among inpatients with CDI. Methods: This retrospective cohort study was conducted at China Medical University Hospital, a 2111-bed tertiary medical center in central Taiwan. We included adult inpatients who had a first positive C. difficile culture or toxin assay and had diarrhea as the study population. The main exposure of interest was the biochemical profiles of white blood cell count, serum creatinine (SCr), estimated glomerular filtration rate, blood urea nitrogen (BUN), serum albumin, and glucose. The primary outcome was the 30-day all-cause mortality and the secondary outcome was the length of stay in the intensive care units (ICU) following CDI. A multivariable Cox model and a logistic regression model were developed using clinically relevant and statistically significant variables for 30-day mortality and for length of ICU stay, respectively. A risk scoring system was established by standardizing the coefficients. We compared the performance of our models and the guidelines. Results: Of 401 patients, 23.4% died within 30 days. In the multivariable model, malignancy (hazard ratio [HR] = 1.95), ≥ 1.5-fold rise in SCr (HR = 2.27), BUN-to-SCr ratio > 20 (HR = 2.04), and increased glucose (≥ 193 vs < 142 mg/dL, HR = 2.18) were significant predictors of 30-day mortality. For patients who survived the first 30 days of CDI, BUN-to-SCr ratio > 20 (Odds ratio [OR] = 4.01) was the only significant predictor for prolonged (> 9 days) length of ICU stay following CDI. The Harrell's c statistic of our Cox model for 30-day mortality (0.727) was significantly superior to those of SHEA-IDSA 2010 (0.645), SHEA-IDSA 2018 (0.591), and ECSMID (0.650). Similarly, the conventional c statistic of our logistic regression model for prolonged ICU stay (0.737) was significantly superior to that of the guidelines (SHEA-IDSA 2010, c = 0.600; SHEA-IDSA 2018, c = 0.634; ESCMID, c = 0.645). Our risk prediction scoring system for 30-day mortality correctly reclassified 20.7, 32.1, and 47.9% of patients, respectively. Conclusions: Our model that included novel biomarkers of BUN-to-SCr ratio and glucose have a higher predictive performance of 30-day mortality and prolonged ICU stay following CDI than do the guidelines.

Variability of red blood cell size predicts all-cause mortality, but not progression to dialysis, in patients with chronic kidney disease: A 13-year pre-ESRD registry-based cohort.

BACKGROUND: Prognostic role of red blood cell distribution width (RDW) in patients with chronic kidney disease (CKD) is unclear. Little evidence provides a comprehensive predictive analysis considering both baseline values and longitudinal trajectories of RDW along with mean corpuscular volume (MCV). METHODS: We conducted a comprehensive risk assessment of RDW and MCV in a registry-based cohort of 4621 patients with CKD (age, 20-90 y) receiving multidisciplinary care during 2003 to 2015. Both baseline and longitudinal trajectories of RDW and MCV were modeled as predictors for end-stage renal disease (ESRD) and mortality by using multiple Cox proportional hazards regression models, incorporating time-varying covariates and adjustments for imperative confounding variables. RESULTS: Fully adjusted hazard ratio (HR; 95% CI) of progression to ESRD for each unit increase in RDW and MCV at baseline was 0.97 (0.93-1.02) and 1.00 (0.99-1.01), respectively. Longitudinally, neither RDW nor MCV trajectory was associated with progression to ESRD. For all-cause mortality, fully adjusted HRs (95%CI) were 1.09 (1.04-1.14) for each percent increase in RDW with a linear dose-response relationship and 1.95 (1.47-2.59) for a stable-high RDW trajectory compared with normal RDW trajectory. The effects of RDW on mortality were further augmented in patients with concomitantly high MCV status. Incorporating point-of-care RDW significantly improves the discrimination performance quantified using Harrell C statistics into the existing CKD mortality predictive equation (from 0.770 to 0.784, P = .018). CONCLUSIONS: We support the clinical utility of RDW in predicting all-cause mortality among patients with CKD. The mechanism underlying our findings is critical for CKD risk assessment and management, particularly from malnutrition, inflammation, and atherosclerosis perspectives.

Automation of the kidney function prediction and classification through ultrasound-based kidney imaging using deep learning.

Prediction of kidney function and chronic kidney disease (CKD) through kidney ultrasound imaging has long been considered desirable in clinical practice because of its safety, convenience, and affordability. However, this highly desirable approach is beyond the capability of human vision. We developed a deep learning approach for automatically determining the estimated glomerular filtration rate (eGFR) and CKD status. We exploited the transfer learning technique, integrating the powerful ResNet model pretrained on an ImageNet dataset in our neural network architecture, to predict kidney function based on 4,505 kidney ultrasound images labeled using eGFRs derived from serum creatinine concentrations. To further extract the information from ultrasound images, we leveraged kidney length annotations to remove the peripheral region of the kidneys and applied various data augmentation schemes to produce additional data with variations. Bootstrap aggregation was also applied to avoid overfitting and improve the model's generalization. Moreover, the kidney function features obtained by our deep neural network were used to identify the CKD status defined by an eGFR of <60 ml/min/1.73 m2. A Pearson correlation coefficient of 0.741 indicated the strong relationship between artificial intelligence (AI)- and creatinine-based GFR estimations. Overall CKD status classification accuracy of our model was 85.6% -higher than that of experienced nephrologists (60.3%-80.1%). Our model is the first fundamental step toward realizing the potential of transforming kidney ultrasound imaging into an effective, real-time, distant screening tool. AI-GFR estimation offers the possibility of noninvasive assessment of kidney function, a key goal of AI-powered functional automation in clinical practice.

Longitudinal lipid trends and adverse outcomes in patients with CKD: a 13-year observational cohort study.

Studies on the effects of longitudinal lipid trajectories on end-stage renal disease (ESRD) development and deaths among patients with chronic kidney disease (CKD) are limited. We conducted a registry-based prospective study using data from a 13-year multidisciplinary pre-ESRD care program. The final study population comprised 4,647 patients with CKD. Using group-based trajectory modeling, we dichotomized longitudinal trajectories of total cholesterol (T-CHO), triglyceride (TG), LDL cholesterol (LDL-C), and HDL cholesterol (HDL-C). Time to ESRD or death was analyzed using multiple Cox regression. At baseline, higher levels of T-CHO and LDL-C were associated with rapid progression to ESRD, whereas only HDL-C was positively associated with all-cause mortality [adjusted hazard ratio (HR), 1.20; 95% CI, 1.06-1.36; P-value, 0.005]. Compared with those with a normal T-CHO trajectory, the fully adjusted HR of patients with a high T-CHO trajectory for ESRD risk was 1.21 (P-value, 0.019). Subgroup analysis showed that a high TG trajectory was associated with a 49% increase in mortality risk in CKD patients without diabetes (P-value for interaction, 0.012). In contrast to what was observed based on baseline HDL-C, patients with a trajectory of frequent hypo-HDL cholesterolemia had higher risk of all-cause mortality (adjusted HR, 1.53; P-value, 0.014). Thus, only T-CHO, both at baseline and over the longitudinal course, demonstrated a significant potential risk of incident ESRD. The inconsistency in the observed directions of association between baseline levels and longitudinal trajectories of HDL-C warrants further research to unveil specific pathogenic mechanisms underlying the HDL-C metabolism in patients with CKD.

First-year estimated glomerular filtration rate variability after pre-end-stage renal disease program enrollment and adverse outcomes of chronic kidney disease.

BACKGROUND: Scarce evidence associates the first-year estimated glomerular filtration rate (eGFR) variability and longitudinal change scales concomitantly to the risk of developing end-stage renal disease (ESRD), acute coronary syndrome (ACS) and death following pre-ESRD program enrollment in chronic kidney disease (CKD). METHODS: We conducted a prospective cohort study of 5092 CKD patients receiving multidisciplinary care between 2003 and 2015 with careful ascertainment of ESRD, ACS and death during the follow-up. First-year eGFR variability and longitudinal change scales that were based on all first-year eGFR measurements included coefficient of variation of eGFR (eGFR-CV), percent change (eGFR-PC), absolute difference (eGFR-AD), slope (eGFR-slope) and area under the curve (AUC). RESULTS: A total of 786 incident ESRD, 292 ACS and 410 death events occurred during the follow-up. In the multiple Cox regression, the fully adjusted hazard ratios (HRs) of progression to ESRD for each unit change in eGFR-CV, eGFR-PC, eGFR-AD, eGFR-slope, eGFR-AUC were 1.03 [95% confidence interval (CI) 1.02-1.04], 1.04 (1.03-1.04), 1.16 (1.14-1.18), 1.16 (1.14-1.17) and 1.04 (1.03-1.04), respectively. The adjusted HRs for incident ESRD comparing the extreme with the reference quartiles of eGFR-CV, eGFR-PC, eGFR-AD, eGFR-slope and eGFR-AUC were 2.67 (95% CI 2.11-3.38), 8.34 (6.33-10.98), 19.08 (11.89-30.62), 13.08 (8.32-20.55) and 6.35 (4.96-8.13), respectively. Similar direction of the effects on the risk of developing ACS and mortality was observed. In the 2 × 2 risk matrices, patients with the highest quartile of eGFR-CV and concomitantly with the most severely declining quartiles of any other longitudinal eGFR change scale had the highest risk of all outcomes. CONCLUSIONS: The dynamics of eGFR changes, both overall variability and longitudinal changes, over the first year following pre-ESRD program enrollment are crucial prognostic factors for the risk of progression to ESRD, ACS and deaths among patients with CKD. A risk matrix combining the first-year eGFR variability and longitudinal change scales following pre-ESRD enrollment is a novel approach for risk characterization in CKD care. Randomized trials in CKD may be required to ascertain comparable baseline eGFR dynamics.

Longitudinal progression trajectory of random urine creatinine as a novel predictor of ESRD among patients with CKD.

BACKGROUND: The clinical importance of random urine creatinine concentration in CKD population remains undetermined. Earlier studies found that lower 24-h urine creatinine excretion was associated with the risk of ESRD and all-cause mortality among CKD patients. METHODS: We modeled the longitudinal trajectories of serial random urine creatinine among 4689 CKD patients enrolled in a national registry-based pre-ESRD program between 2003 and 2015 at a tertiary medical center. Other biochemical parameters including kidney function and serum albumin were regularly evaluated. Primary study outcomes were ESRD requiring maintenance dialysis and all-cause mortality. RESULTS: By group-based trajectory modeling, the urine creatinine trajectories were characterized into three patterns: (1) stable low; (2) medium; and (3) high-declining. The adjusted hazard ratio of incident ESRD and all-cause mortality increased by 6% (95% CI: 1-12%) and 9% (95% CI: 2-17%), respectively, for each 20 mg/dL reduction in baseline random urine creatinine concentration. Consistently, there was a significant inverse linear dose-response relationship between baseline random urine creatinine and incident ESRD, but not all-cause mortality. Compared to patients with "medium" and "high-declining" urine creatinine trajectories combined, the adjusted hazard ratio for incidental ESRD among patients with a "stable-low" trajectory who had serial random urine creatinine concentrations stably below 100 mg/dL was 1.46 (95% CI: 1.00-2.12) after considering the competing risk of death. CONCLUSIONS: Random urine creatinine not only serves as a common urinary concentration corrector but has its own clinical significance in risk stratification and outcome prediction in patients with advanced CKD.