Machine learning model predicts clotting risk during CRRT in ESKD patients: a SHAP-interpretable approach

Article Figures and data Abstract eLife digest Introduction Results Discussion Materials and methods Data availability References Decision letter Author response Article and author information Metrics Abstract End-stage kidney disease (ESKD) patients are at high risk of severe COVID-19. We measured 436 circulating proteins in serial blood samples from hospitalised and non-hospitalised ESKD patients with COVID-19 (n = 256 samples from 55 patients). Comparison to 51 non-infected patients revealed 221 differentially expressed proteins, with consistent results in a separate subcohort of 46 COVID-19 patients. Two hundred and three proteins were associated with clinical severity, including IL6, markers of monocyte recruitment (e.g. CCL2, CCL7), neutrophil activation (e.g. proteinase-3), and epithelial injury (e.g. KRT19). Machine-learning identified predictors of severity including IL18BP, CTSD, GDF15, and KRT19. Survival analysis with joint models revealed 69 predictors of death. Longitudinal modelling with linear mixed models uncovered 32 proteins displaying different temporal profiles in severe versus non-severe disease, including integrins and adhesion molecules. These data implicate epithelial damage, innate immune activation, and leucocyte–endothelial interactions in the pathology of severe COVID-19 and provide a resource for identifying drug targets. eLife digest COVID-19 varies from a mild illness in some people to fatal disease in others. Patients with severe disease tend to be older and have underlying medical problems. People with kidney failure have a particularly high risk of developing severe or fatal COVID-19. Patients with severe COVID-19 have high levels of inflammation, causing damage to tissues around the body. Many drugs that target inflammation have already been developed for other diseases. Therefore, to repurpose existing drugs or design new treatments, it is important to determine which proteins drive inflammation in COVID-19. Here, Gisby, Clarke, Medjeral-Thomas et al. measured 436 proteins in the blood of patients with kidney failure and compared the levels between patients who had COVID-19 to those who did not. This revealed that patients with COVID-19 had increased levels of hundreds of proteins involved in inflammation and tissue injury. Using a combination of statistical and machine learning analyses, Gisby et al. probed the data for proteins that might predict a more severe disease progression. In total, over 200 proteins were linked to disease severity, and 69 with increased risk of death. Tracking how levels of blood proteins changed over time revealed further differences between mild and severe disease. Comparing this data with a similar study of COVID-19 in people without kidney failure showed many similarities. This suggests that the findings may apply to COVID-19 patients more generally. Identifying the proteins that are a cause of severe COVID-19 – rather than just correlated with it – is an important next step that could help to select new drugs for severe COVID-19. Introduction Coronavirus disease (COVID-19), caused by the SARS-CoV-2 virus, displays wide clinical heterogeneity from asymptomatic to fatal disease. Patients with severe disease exhibit marked inflammatory responses and immunopathology. The mechanisms underlying this remain incompletely characterised, and the key molecular mediators are yet to be determined. The first treatment shown to reduce mortality from COVID-19 in randomised trials was dexamethasone (Horby et al., 2020), a corticosteroid that has broad non-specific effects on the immune system. Even with corticosteroid treatment, mortality in severe COVID-19 remains significant. There is a wide armamentarium of existing drugs that target inflammation more selectively, providing potential repurposing opportunities for the treatment of COVID-19. Recently, the REMAP-CAP trial has demonstrated efficacy of anti-IL6 receptor blockade in patients admitted to intensive care units with severe disease (Gordon et al., 2021). In order to select the most promising agents for future trials, we urgently need to better understand the molecular drivers of severe disease. Proteins are the effector molecules of biology and the targets of most drugs. Therefore, proteomic profiling to identify the key mediators of severe disease provides a valuable tool for identifying and prioritising potential drug targets (Suhre et al., 2021). Risk factors for severe or fatal COVID-19 include age, male sex, non-European ancestry, obesity, diabetes mellitus, cardiovascular disease, and immunosuppression (Williamson et al., 2020). End-stage kidney disease (ESKD) is one of the strongest risk factors for severe COVID-19 (estimated hazard ratio for death 3.69) (Williamson et al., 2020), and ESKD patients hospitalised with COVID-19 have a mortality of approximately 30% (Docherty et al., 2020; Corbett et al., 2020; Ng et al., 2020; Valeri et al., 2020). ESKD patients have a high prevalence of vascular and cardiometabolic disease (e.g. hypertension, ischaemic heart disease, diabetes), either as a result of the underlying cause of their renal disease or as a consequence of renal failure. In addition, ESKD results in both relative immunosuppression and chronic low-grade inflammation, which may impact viral defence and the host inflammatory response. Here we performed proteomic profiling of serial blood samples of ESKD patients with COVID-19, leveraging the unique opportunity for longitudinal sampling in both the outpatient and inpatient settings afforded by a large multi-ethnic haemodialysis cohort (Figure 1a). These data revealed 221 proteins that are dysregulated in COVID-19 versus matched non-infected ESKD patients. Using linear mixed models, joint models, and machine learning, we identified proteins that are markers of COVID-19 severity and risk of death. Finally, we characterised the temporal dynamics of the blood proteomic response during COVID-19 infection in ESKD patients, uncovering 32 proteins that display altered trajectories in patients with severe versus non-severe disease. Figure 1 with 1 supplement see all Download asset Open asset Study design. (a) Schematic representing a summary of the patient cohorts, sampling, and the major analyses. Blue and red stick figures represent outpatients and hospitalised patients, respectively. (b) Timing of serial blood sampling in relation to clinical course of COVID-19 (subcohort A). Black asterisks indicate when samples were obtained. Three patients were already in hospital prior to COVID-19 diagnosis (indicated by red bars). Results We recruited 55 ESKD patients with COVID-19 (subcohort A; Table 1). All patients were receiving haemodialysis prior to acquiring COVID-19. Blood samples were taken as soon as feasible following COVID-19 diagnosis. At time of initial sample, 30 patients were outpatients attending haemodialysis sessions and 25 were hospitalised inpatients (see Materials and methods, Figure 1). Following the initial blood sample, serial sampling was performed for 51/55 patients. We also recruited 51 non-infected haemodialysis patients as ESKD controls, mirroring the age, sex, and ethnicity distribution of the COVID-19 cases (Figure 1—figure supplement 1a–c). We used the Olink proteomics platform to measure 436 proteins (Supplementary file 1a) in 256 plasma samples from the COVID-19 patients and the 51 control samples. The proteins measured consisted of five multiplex 'panels' focussed on proteins relevant to immuno-inflammation, cardiovascular, and cardiometabolic disease. The 436 proteins assayed showed strong enrichment for immune-related proteins (Supplementary file 1b). Table 1 Characteristics of subcohort A. COVID-19-positive ESKD patients (n = 55)ESKD controls (n = 51) OverallPeak severity mild or moderate (n = 28)Peak severity severe or critical (n = 27)Age Median (IQR)72.2 62.5–77.373.4 65.5–76.468.5 61.8–78.870.1 62.2–75.1Sex M F39 (70.9%) 16 (29.1%)18 (64.3%) 10 (35.7%)21 (77.8%) 6 (22.2%)36 (70.6%) 15 (29.4%)Ethnicity White Black South Asian Asian (other) Other16 (29.1%) 8 (14.5%) 18 (32.7%) 4 (7.3%) 9 (16.4%)5 (17.9%) 5 (17.9%) 10 (35.7%) 1 (3.6%) 7 (25.0%)11 (40.7%) 3 (11.1%) 8 (29.6%) 3 (11.1%) 2 (7.4%)13 (25.5%) 8 (15.7%) 20 (39.2%) 3 (5.9%) 7 (13.7%)Diabetes34 (61.8%)*16 (57.1%)18 (66.7%)24 (47.1%)*Current smoker1 (1.8%)1 (3.6%)00ESKD cause DN Genetic GN HTN/vascular Other Unknown29 (52.7%) 1 (1.8%) 3 (5.5%) 5 (9.1%) 8 (14.5%) 9 (16.4%)14 (50.0%) 1 (3.6%) 1 (3.6%) 3 (10.7%) 5 (17.9%) 4 (14.3%)15 (55.6%) 0 2 (7.4%) 2 (7.4%) 3 (11.1%) 5 (18.5%)20 (39.2%) 1 (2.0%) 9 (17.6%) 7 (13.7%) 4 (7.8%) 10 (19.6%)Hospitalisation due to COVID-19†33 (60%)6 (21.4%)27 (100%)N/AFatal COVID-199 (16.3%)0 (0%)9 (33.3%)N/A DN = diabetic nephropathy. GN = glomerulonephritis. HTN = hypertension. IQR = inter-quartile range. 'South Asian' represents individuals with Indian, Pakistani, or Bangladeshi ancestry. Subsets defined according to peak WHO severity over the course of the illness. N/A = not applicable. *One patient had type 1 diabetes, the remainder type 2. †3 patients were hospitalised prior to COVID-19 diagnosis. 8 patients diagnosed with COVID-19 as outpatients subsequently deteriorated were hospitalised. In addition, we performed the Olink proteomic assays in 52 serum samples from a separate set of 46 COVID-19-positive ESKD patients (subcohort B) and 11 serum samples from ESKD COVID-19-negative controls (a subset of the controls described above). For the large majority of patients in subcohort B, only a single timepoint was available. A higher proportion of these patients (41/46, 89%) were hospitalised and had severe disease (Table 2) than in subcohort A (Figure 1, Table 1). Table 2 Characteristics of subcohort B. COVID-19-positive ESKD patients (n = 46)COVID-19-negative ESKD controls (n = 11)*Age Median (IQR)64.3 60.3–73.071.6 (61.7–73.9)Sex M F32 (69.6%) 14 (30.4%)8 (72.3%) 3 (27.3%)Ethnicity White Black South Asian Asian (other) Other11 (23.9%) 8 (17.4%) 12 (26.1%) 7 (15.2%) 8 (17.4%)3 (27.3%) 3 (27.3%) 3 (27.3%) 0 2 (18.2%)Diabetes29 (63.0%)6 (54.5%)Current smoker2 (4.3%)0 (%)ESKD cause DN Genetic GN HTN/vascular Other Unknown19 (41.3%) 1 (2.2%) 7 (15.2%) 3 (6.5%) 3 (6.5%) 13 (28.3%)5 (45.5%) 0 1 (9.1%) 1 (9.1%) 2 (18.2%) 2 (18.2%)Hospitalisation due to COVID-1941 (89.1%)N/ASevere or critical COVID-1933 (71.7%)N/AFatal COVID-199 (19.6%)N/A DN = diabetic nephropathy. GN = glomerulonephritis. HTN = hypertension. IQR = inter-quartile range. 'South Asian' represents individuals with Indian, Pakistani, or Bangladeshi ancestry. Subsets defined according to peak WHO severity over the course of the illness. N/A = not applicable. *These 11 controls are a subset of the control patients used in subcohort A. Proteomic differences between COVID-19-positive and -negative ESKD patients Principal component analysis (PCA) of proteomic data from subcohort A demonstrated differences between samples from COVID-19-positive cases and controls, although the two groups did not separate into discrete clusters (Figure 2a,b). To examine the effects of COVID-19 on the plasma proteome, we performed a differential expression analysis in subcohort A between COVID-19 cases (n = 256 samples passing quality control [QC] from 55 patients) and non-infected ESKD controls (n = 51) using linear mixed models, which account for serial samples from the same individual (see Materials and methods). This revealed 221 proteins associated with COVID-19 (5% false discovery rate, FDR); the vast majority were upregulated, with only 40 downregulated (Figure 3a, Supplementary file 1c). In order to provide a succinct and standardised nomenclature, we report proteins by the symbols of the genes encoding them (see Supplementary file 1a for a mapping of symbols to full protein names). The most strongly upregulated proteins (in terms of fold change) were DDX58, CCL7, IL6, CXCL11, KRT19, and CXCL10, and the most strongly downregulated were SERPINA5, CCL16, FABP2, PON3, ITGA11, and MMP12 (Figure 3—figure supplement 1). Notably, many of the upregulated proteins were chemotaxins. Figure 2 with 2 supplements see all Download asset Open asset Principal component analysis. PC = principal component. Each point represents a sample. Colouring indicates COVID-19 status. The directions and relative sizes of the six largest PC loadings are plotted as arrows (middle column). (a, b) Subcohort A. Due to serial sampling, there are multiple samples for most patients. The proportion of variance explained in subcohort A by each PC is shown in parentheses on the axis labels. (c, d) Subcohort B. Samples are projected into the PCA coordinates from subcohort A. Figure 3 with 4 supplements see all Download asset Open asset Identification of dysregulated proteins. (a) Proteins upregulated (red) or downregulated (blue) in COVID-19-positive patients versus COVID-19-negative ESKD patients n = 256 plasma samples from 55 COVID-19-positive patients, versus n = 51 ESKD controls (one sample per control patient). (b) Proteins associated with disease severity associations of protein levels against WHO severity score at the time of sampling. Linear gradient indicates the effect size. A positive effect size (red) indicates that an increase in protein level is associated with increasing disease severity and a negative gradient (blue) the opposite. n = 256 plasma samples from 55 COVID-19-positive patients. For (a, b), p-values from linear mixed models after Benjamini–Hochberg adjustment; significance threshold = 5% FDR; dark-grey = non-significant. (c) Heatmap showing protein levels for selected proteins with strong associations with severity. Each column represents a sample (n = 256 COVID-19 samples and 51 non-infected samples). Each row represents a protein. Proteins are annotated using the symbol of their encoding gene. For the purposes of legibility, not all significantly associated proteins are shown; the heatmap is limited to the 17% most up- or downregulated proteins (by effect size) of those with a significant association. Proteins are ordered by hierarchical clustering. Samples are ordered by WHO severity at the time of blood sample ('Severity'). 'Overall course' indicates the peak WHO severity over the course of the illness. We observed that a high proportion of the measured proteins were associated with COVID-19. Given the highly targeted nature of the Olink panels that we used (enriched for immune and inflammation-related proteins), this was not surprising. Nevertheless, to ensure that the Benjamini–Hochberg adjustment of p-values was controlling the FDR at the 5% level, we performed two additional analyses (see Materials and methods). First, we estimated the FDR using an alternative method (the plug-in procedure ; Hastie et al., 2001); this confirmed appropriate FDR control. Second, we used permutation to estimate the distribution of the number of proteins expected to be declared significant under the null hypothesis of no association between any proteins and COVID-19. This showed that the probability of observing the number of differentially abundant proteins we identified was highly unlikely under the null (empirical p<1×10−5; Figure 3—figure supplement 2). Although our COVID-19-negative controls were well matched in terms of age, sex, and ethnicity (Figure 1—figure supplement 1a–c), perfect matching of comorbidities was not feasible in the context of the healthcare emergency at the time of patient recruitment. There was a higher prevalence of diabetes in the COVID-19 cases compared to the controls (61.8% versus 47.1%, respectively; Table 1). To evaluate whether differing rates of diabetes had impacted the proteins identified as differentially abundant between cases and controls, we performed a sensitivity analysis adding diabetes as an additional covariate in the linear mixed model. This did not materially affect our findings; estimated effect sizes and –log10 p-values from models with and without the inclusion of diabetes were highly correlated (Pearson r > 0.99, and r = 0.95, respectively; Figure 3—figure supplement 3a,b). Full results from both models are shown in Supplementary file 1c. Similarly, there were also differences in the underlying cause of ESKD in cases compared to controls (Table 1). We therefore performed a further sensitivity analysis adjusting for underlying cause of renal failure. This did not make any meaningful difference to our results (Figure 3—figure supplement 3c,d, Supplementary file 1c). We also considered the possibility that timing of haemodialysis might affect the plasma proteome. To minimise the impact of this, all samples were taken prior to haemodialysis. For the large majority (86.6%) of samples, the most recent haemodialysis was between 48 and 72 hr prior to blood draw. This consistency in timing of blood sampling reduces the potential for impact of this issue. Nevertheless, to evaluate whether timing of haemodialysis might have impacted our results, we performed a sensitivity analysis including time from last haemodialysis as a covariate. Our results were not materially affected by this, with −log10 p-values and estimated effect sizes very highly correlated with those obtained without inclusion of this covariate (Pearson r > 0.99 for effect size estimates and for −log10 p-values; Figure 3—figure supplement 4a,b, Supplementary file 1c). We used the smaller subcohort B (n = 52 serum samples from 46 patients with COVID-19; see Materials and methods) for validation. We first projected the data from subcohort B into the PCA space of subcohort A to examine how well the separation of cases and controls in the PCA space replicated (see Materials and methods). This revealed clearer separation of infected and non-infected patients than in subcohort A (Figure 2c,d), perhaps reflecting the higher proportion of hospitalised patients (41 of 46 patients) in subcohort B (Table 2). We next performed differential abundance analysis in subcohort B and found 201 proteins that were dysregulated in cases versus controls (5% FDR) (Supplementary file 1c). Of the 221 differentially abundant proteins from subcohort A, 150 (69.7%) were also identified in subcohort B at 5% FDR (Figure 4a). Effect sizes in each dataset showed a strong correlation (r = 0.80, Figure 4b). This demonstrates that our findings are highly reproducible despite differences in sample sizes and blood materials (plasma versus serum in subcohorts A and B, respectively). Figure 4 with 1 supplement see all Download asset Open asset Validation. (a) Overlap between the significant associations in the differential abundance analysis between ESKD patients with and without COVID-19 in subcohorts A and B. 5% FDR was used as the significance threshold in both analyses. (b) Comparison of estimated effect sizes for all 436 proteins in the differential abundance analyses (COVID-19 positive versus negative) in subcohort A and B. Each point represents a protein. Pearson's r is shown. Differential abundance analyses were performed using linear mixed models. Subcohort A analysis (plasma samples): 256 samples from 55 COVID-19 patients versus 51 non-infected patient samples (single time-point). Subcohort B (serum samples): 52 samples from 55 COVID-19 patients and 11 non-infected patient samples (single timepoint). Proteins associated with COVID-19 severity Examination of the principal components plot labelling samples by clinical severity at the time of sampling (defined by WHO severity scores, graded as mild, moderate, severe, or critical) demonstrated a gradient of COVID-19 severity, best captured by principal components 1 and 3 (Figure 2—figure supplement 1a). To determine the proteomic effects of COVID-19 severity, we tested for associations between proteins and WHO severity score at the time of blood sampling, using linear mixed models with severity encoded as an ordinal predictor (see Materials and methods). This analysis revealed 203 proteins associated with severity (Figure 3b, Supplementary file 1d). The majority of these were upregulated in more severe disease, with only 42 downregulated. A sensitivity analysis adjusting for time since last haemodialysis made no significant impact on our results (Figure 3—figure supplement 4c,d, Supplementary file 1d). Consistent with previous reports, we found that severe COVID-19 was characterised by elevated IL6. In addition, we observed a signature of upregulated monocyte chemokines (e.g. CCL2, CCL7, CXCL10), neutrophil activation and degranulation (e.g. PRTN3, MPO), and epithelial injury (e.g. KRT19, AREG, PSIP1, GRN). (Figures 3b,c and 5). SERPINA5 and leptin showed the greatest downregulation as COVID-19 severity increased (Figure 3b,c). Figure 5 Download asset Open asset proteins strongly associated with COVID-19 severity. showing distribution of plasma protein levels according to COVID-19 at the time of blood draw. indicate and inter-quartile range. n = 256 samples from 55 COVID-19 patients and 51 samples from non-infected patients. WHO severity indicates the clinical severity score of the patient at the time the sample was n = moderate n = severe n = critical n = 15 samples. monocyte markers of epithelial injury. two neutrophil and IL6. We next how the COVID-19 severity protein signature to the proteins that are differentially abundant between cases and The majority of proteins were also identified as differentially abundant in the COVID-19-positive versus -negative analysis (Figure fold for proteins in COVID-19 versus non-infected patients were correlated with effect sizes in the severity that the proteins most upregulated in cases versus controls also to the greatest in severe disease (Figure there were some (e.g. that were strongly associated with severity, not differentially expressed in infected versus non-infected patients (Figure Figure 6 Download asset Open asset Comparison of proteins differentially expressed in COVID-19 with those associated with clinical severity. (a) Overlap between the proteins significantly differentially expressed in COVID-19 (n = 256 COVID-19 samples and 51 non-infected versus those associated with severity n = 256 (subcohort A). 5% FDR was used as the significant in both analyses. (b) Comparison of effect sizes for each protein in the COVID-19-positive versus -negative analysis and severity analysis Each point represents a protein. Pearson's r is shown. (c) of proteins associated with severity, not significantly differentially abundant in the of all cases versus showing distribution of plasma protein levels according to COVID-19 at the time of blood draw. indicate and inter-quartile range. n = 256 samples from 55 COVID-19 patients and 51 samples from non-infected patients. WHO severity indicates the clinical severity score of the patient at the time the sample was n = moderate n = severe n = critical n = 15 samples. learning to predict COVID-19 severity PCA revealed that some samples from patients who had mild or moderate disease at the time of sampling with samples from patients with severe disease (Figure 2—figure supplement 1a). Examination of the same PCA plot labelling samples according to the clinical course by peak WHO severity score over the of the (Figure 2—figure supplement revealed that these samples from individuals who subsequently developed severe or critical disease. This that molecular may clinical To evaluate this we used learning to whether the proteomic signature of the first blood sample for each patient in our dataset could identify whether the patient either had severe COVID-19 at the time of sampling or severe disease in the differential expression analyses each protein of all proteins in the Using we a on the first sample for each COVID-19 patient to predict the clinical defined by peak WHO severity. For the purposes of this we clinical course into either WHO or The method in peak severity. using only predictors and clinical the method in peak severity. clinical proteins did not compared to using proteomic predictors that the information in the clinical predictors is captured at the proteomic we not that proteomic profiling is to clinical for risk during this the selected by the proteins of We therefore the to identify key proteins by (see Materials and methods, Supplementary file The most important proteins for the of or future severe disease were IL18BP, GDF15, KRT19, and (Figure is that this as a key of severe disease. Figure 7 with 1 supplement see all Download asset Open asset of severe COVID-19 and death. (a) The 12 most important proteins for clinical course (defined by peak COVID-19 WHO using a is important for it is to in many of and be to the have a The all was used as the (b) Proteins that are significant predictors of death n = 256 samples from 55 COVID-19-positive patients, of Risk estimates are from a joint model. indicate For proteins with a positive risk a higher to a high risk of and for proteins with negative Proteins associated with risk of death of 55 patients in subcohort A We therefore to identify proteins associated with risk of death. To the nature of protein for of we joint models, which linear mixed models and models et al., (see Materials and methods). This analysis identified proteins for which increased was associated with increased risk of death (Figure Supplementary file including and and 25 proteins for which increased was associated with risk of including and with clinical A number of clinical have associations with COVID-19 (e.g. elevated inflammatory and et al., 2020). We therefore compared our proteomic data from COVID-19 patients at each timepoint to clinical using linear mixed models (see Materials and methods). We found associations between plasma proteins and all clinical (Figure Supplementary file Many of these proteins were also markers of severity (e.g. IL6, KRT19, and were strongly associated with and and Of CCL7, a monocyte that was also identified as an important of severity by the was associated with monocyte and inflammatory neutrophil was associated with which IL6, and and with the and Figure 8 Download asset Open asset of clinical markers with plasma proteins. Proteins that are (red) or (blue) associated with clinical (5% p-values from differential abundance analysis using linear mixed models after Benjamini–Hochberg = non-significant. Two associations were found for shown – see Supplementary file Longitudinal analysis proteins with temporal profiles according to severity The immune response to infection is and therefore provide only the serial sampling in our dataset (Figure we the temporal of each protein and whether or not any protein trajectories in patients with a versus clinical This was a

Undernutrition among children under the age of five is a major public health concern, especially in developing countries. This study aimed to use machine learning (ML) algorithms to predict undernutrition and identify its associated factors. Secondary data analysis of the 2017 Multiple Indicator Cluster Survey (MICS) was performed using R and Python. The main outcomes of interest were undernutrition (stunting: height-for-age (HAZ) < -2 SD; wasting: weight-for-height (WHZ) < -2 SD; and underweight: weight-for-age (WAZ) < -2 SD). Seven ML algorithms were trained and tested: linear discriminant analysis (LDA), logistic model, support vector machine (SVM), random forest (RF), least absolute shrinkage and selection operator (LASSO), ridge regression, and extreme gradient boosting (XGBoost). The ML models were evaluated using the accuracy, confusion matrix, and area under the curve (AUC) receiver operating characteristics (ROC). In total, 8564 children were included in the final analysis. The average age of the children was 926 days, and the majority were females. The weighted prevalence rates of stunting, wasting, and underweight were 17%, 7%, and 12%, respectively. The accuracies of all the ML models for wasting were (LDA: 84%; Logistic: 95%; SVM: 92%; RF: 94%; LASSO: 96%; Ridge: 84%, XGBoost: 98%), stunting (LDA: 86%; Logistic: 86%; SVM: 98%; RF: 88%; LASSO: 86%; Ridge: 86%, XGBoost: 98%), and for underweight were (LDA: 90%; Logistic: 92%; SVM: 98%; RF: 89%; LASSO: 92%; Ridge: 88%, XGBoost: 98%). The AUC values of the wasting models were (LDA: 99%; Logistic: 100%; SVM: 72%; RF: 94%; LASSO: 99%; Ridge: 59%, XGBoost: 100%), for stunting were (LDA: 89%; Logistic: 90%; SVM: 100%; RF: 92%; LASSO: 90%; Ridge: 89%, XGBoost: 100%), and for underweight were (LDA: 95%; Logistic: 96%; SVM: 100%; RF: 94%; LASSO: 96%; Ridge: 82%, XGBoost: 82%). Age, weight, length/height, sex, region of residence and ethnicity were important predictors of wasting, stunting and underweight. The XGBoost model was the best model for predicting wasting, stunting, and underweight. The findings showed that different ML algorithms could be useful for predicting undernutrition and identifying important predictors for targeted interventions among children under five years in Ghana.

This study aimed to utilize various machine learning algorithms to develop a predictive model for the progression of severe community-acquired pneumonia (SCAP) in children to critical severe community-acquired pneumonia (cSCAP). Retrospective analysis of clinical data of SCAP patients admitted to the Department of Pediatric Intensive Care Medicine at the First Affiliated Hospital of Bengbu Medical University from January 2021 to April 2023. Logistic regression (LR) and Least Absolute Shrinkage and Selection Operator (LASSO) were jointly employed to screen model variables. The selected variables were then incorporated into seven algorithms, namely LR, Decision Tree (DT), Random Forest (RF), Extreme Gradient Boosting (XGBoost), Naive Bayes (NB), k-Nearest Neighbor (KNN), and Support Vector Machine (SVM), to establish a predictive model for the progression of SCAP in children to a critically severe stage. The effectiveness of the model was evaluated based on the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and F1 score. Finally, the Shapley Additive Explanation (SHAP) algorithm was used to interpret the established machine learning model. A total of 211 patients were included. Red Cell Distribution Width–Coefficient of Variation (RDW-CV), procalcitonin (PCT), blood urea nitrogen (BUN), and lactate dehydrogenase (LDH) were selected as predictors. The XGBoost model outperformed six other algorithms, with an AUC of 0.98 (95% CI,0.93–1.00 ), accuracy 0.89 (95% CI, 0.78–0.94), sensitivity 0.98 (95% CI, 0.95–1.00), and specificity 0.75 (95% CI, 0.45–0.87). SHAP analysis identified PCT, LDH, RDW-CV, and BUN as the most important contributors, supporting their clinical relevance for early risk stratification. This study developed an accurate predictive model for the cSCAP in children using machine learning techniques, providing clinical support for decision-making by clinicians.

Accurate prediction of in-hospital mortality for patients with severe community-acquired pneumonia (SCAP) complicated by respiratory failure admitted to the intensive care unit (ICU) remains a critical challenge. This study aimed to develop and validate a machine learning (ML) model to predict this risk and compare its performance with conventional scoring systems. In this retrospective study, data from 164 patients with SCAP and respiratory failure admitted to the ICU between January 2017 and January 2024 were analyzed. Patients were randomly divided into a training set (n = 114) and a validation (test) set (n = 50). Forty-five clinical features collected at admission were used as candidate predictors. The Least Absolute Shrinkage and Selection Operator (LASSO) regression was employed for feature selection. Six ML models, including Gradient Boosting Decision Tree (GBDT), Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Decision Tree(DT),Support Vector Machine (SVM), and Logistic Regression(LR), were constructed and evaluated.Use SHAP analysis to assess the contribution of each feature in a machine learning model. Construct a nomogram using the top six most influential features. The GBDT model demonstrated the best predictive performance, achieving an area under the receiver operating characteristic curve (AUC) of 0.83 (95% CI: 0.757-0.927) in the internal validation set, significantly outperforming the Acute Physiology and Chronic Health Evaluation II (APACHE-II, AUC = 0.70). Calibration curves demonstrated good agreement between predicted and observed mortality risks, particularly across the mid-probability range. Decision curve analysis indicated that the model provided a higher net benefit than "treat-all" and "treat-none" strategies across a broad range of threshold probabilities. SHapley Additive exPlanations (SHAP) analysis identified lactate, D-dimer, temperature, albumin, Prothrombin Time and Fraction of Inspired Oxygen as the six most influential predictors of in-hospital mortality. Based on these key predictors, we further developed a simplified nomogram to facilitate bedside risk estimation. The GBDT ML model, developed from routinely available clinical data, provides a highly accurate and clinically interpretable tool for predicting in-hospital mortality in SCAP patients with respiratory failure. It outperforms traditional severity scores and holds promise for assisting clinicians in risk stratification and early intervention.

This study sought to develop and validate different machine learning (ML) models that leverage non-contrast MRI radiomics to predict the degree of nonperfusion volume ratio (NVPR) of high-intensity focused ultrasound (HIFU) treatment for uterine fibroids, equipping clinicians with an early prediction tool for decision-making. This study conducted a retrospective analysis on 221 patients with uterine fibroids who received HIFU treatment and were divided into a training set (N = 117), internal validation (N = 49), and an external test set (N = 55). The 851 radiomics features were extracted from T2-weighted imaging (T2WI), and the max-relevance and min-redundancy (mRMR) and the least absolute shrinkage and selection operator (LASSO) regression were applied for feature selection. Several ML models were constructed by logistic regression (LR), decision tree (DT), random forest (RF), support vector machine (SVM), extreme gradient boosting (XGBoost), and light gradient boosting machine (LGBM). These models underwent internal and external validation, and the best model's feature significance was assessed via the Shapley additive explanations (SHAP) method. Four significant non-contrast MRI radiomics features were identified, with the SVM model outperforming others in both internal and external validations, and the AUCs of the T2WI models were 0.860, 0.847, and 0.777, respectively. SHAP analysis highlighted five critical predictors of postoperative NVPR degree, encompassing two radiomics features from non-contrast MRI and three clinical data indicators. The SVM model combining radiomics features and clinical parameters effectively predicts NVPR degree post-HIFU, which enables timely and effective interventions of HIFU.

BackgroundEarly identification of Alzheimer’s disease or mild cognitive impairment can help guide direct prevention and supportive treatments, improve outcomes, and reduce medical costs. Existing advanced diagnostic tools are mostly based on neuroimaging and suffer from certain problems in cost, reliability, repeatability, accessibility, ease of use, and clinical integration. To address these problems, we developed, evaluated, and implemented an early diagnostic tool using machine learning and non-imaging factors.Methods and resultsA total of 654 participants aged 65 or older from the Nursing Home in Hangzhou, China were identified. Information collected from these patients includes dementia status and 70 demographic, cognitive, socioeconomic, and clinical features. Logistic regression, support vector machine (SVM), neural network, random forest, extreme gradient boosting (XGBoost), least absolute shrinkage and selection operator (LASSO), and best subset models were trained, tuned, and internally validated using a novel double cross validation algorithm and multiple evaluation metrics. The trained models were also compared and externally validated using a separate dataset with 1,100 participants from four communities in Zhejiang Province, China. The model with the best performance was then identified and implemented online with a friendly user interface. For the nursing dataset, the top three models are the neural network (AUROC = 0.9435), XGBoost (AUROC = 0.9398), and SVM with the polynomial kernel (AUROC = 0.9213). With the community dataset, the best three models are the random forest (AUROC = 0.9259), SVM with linear kernel (AUROC = 0.9282), and SVM with polynomial kernel (AUROC = 0.9213). The F1 scores and area under the precision-recall curve showed that the SVMs, neural network, and random forest were robust on the unbalanced community dataset. Overall the SVM with the polynomial kernel was found to be the best model. The LASSO and best subset models identified 17 features most relevant to dementia prediction, mostly from cognitive test results and socioeconomic characteristics.ConclusionOur non-imaging-based diagnostic tool can effectively predict dementia outcomes. The tool can be conveniently incorporated into clinical practice. Its online implementation allows zero barriers to its use, which enhances the disease’s diagnosis, improves the quality of care, and reduces costs.

Anti-tuberculosis drug-induced liver injury (ATLI) is the most harmful to anti-tuberculosis (TB) treatment. This study aims to construct and validate a binary ATLI risk prediction model based on clinical data through seven machine learning algorithms (logistic regression, decision tree, support vector machine, random forest, gradient boosting decision tree, extreme gradient boosting, and light gradient boosting machine [LightGBM]). A retrospective cohort of 2356 TB patients followed between January 2017 and December 2024 was used to develop and evaluate the prediction model. Random undersampling was performed to address class imbalance problem. Least absolute shrinkage and selection operator (LASSO) regression was used to select features and retained 27 of 31 original features. Seven ML algorithms were trained and the LightGBM model demonstrated optimal performance in testing set based on the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity (AUC = 0.789, sensitivity = 0.734, and specificity = 0.706). The model exhibited optimal simplicity and stability when incorporating the 8-feature combination comprising baseline high-density lipoprotein cholesterol (HDLC), γ-glutamyl transpeptidase (GGT), triglycerides, total cholesterol (TCHOL), uric acid, total bilirubin (TBIL), globulin (GLB), and liver disease history (AUC = 0.764, sensitivity = 0.758, and specificity = 0.610), and Shapely additive explanations analysis also revealed that these variables were the most influential contributors. The optimal model maintained robust predictive ability in the external validation cohort (AUC = 0.721, sensitivity = 0.828, and specificity = 0.604). This study determined that the combination of baseline HDLC, GGT, triglycerides, total cholesterol (TCHOL), uric acid, TBIL, GLB, and liver disease history was an important predictor of ATLI through LightGBM model, which could help clinicians in the early identification of ATLI.

Mild cognitive impairment (MCI) is a clinical condition characterized by a decline in cognitive ability and progression of cognitive impairment. It is often considered a transitional stage between normal aging and Alzheimer's disease (AD). This study aimed to compare deep learning (DL) and traditional machine learning (ML) methods in predicting MCI using plasma proteomic biomarkers. A total of 239 adults were selected from the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort along with a pool of 146 plasma proteomic biomarkers. We evaluated seven traditional ML models (support vector machines (SVMs), logistic regression (LR), naïve Bayes (NB), random forest (RF), k-nearest neighbor (KNN), gradient boosting machine (GBM), and extreme gradient boosting (XGBoost)) and six variations of a deep neural network (DNN) model-the DL model in the H2O package. Least Absolute Shrinkage and Selection Operator (LASSO) selected 35 proteomic biomarkers from the pool. Based on grid search, the DNN model with an activation function of "Rectifier With Dropout" with 2 layers and 32 of 35 selected proteomic biomarkers revealed the best model with the highest accuracy of 0.995 and an F1 Score of 0.996, while among seven traditional ML methods, XGBoost was the best with an accuracy of 0.986 and an F1 Score of 0.985. Several biomarkers were correlated with the APOE-ε4 genotype, polygenic hazard score (PHS), and three clinical cerebrospinal fluid biomarkers (Aβ42, tTau, and pTau). Bioinformatics analysis using Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) revealed several molecular functions and pathways associated with the selected biomarkers, including cytokine-cytokine receptor interaction, cholesterol metabolism, and regulation of lipid localization. The results showed that the DL model may represent a promising tool in the prediction of MCI. These plasma proteomic biomarkers may help with early diagnosis, prognostic risk stratification, and early treatment interventions for individuals at risk for MCI.

Background and objectiveThreatened abortion, a common pregnancy complication that often leading to abortion, is hard to predict due to its non-specific symptoms and difficulty in differentiating from other early pregnancy bleeding causes. Current diagnostic methods like serial ultrasounds and clinical monitoring are time-consuming and lack timeliness. To fill the gap in using advanced analytics for early detection and risk stratification, this study develops a machine learning (ML) model based on routine blood data to better predict threatened abortion, providing a reference for early detection and intervention.MethodsIn this study, we collected medical records from January 2022 to March 2024. We analyzed data from 1764 patients with threatened abortion and 1489 healthy controls. Blood test data of all participants were gathered. The Z-score normalization technique was applied to standardize blood routine indicators. This reduced the influence of outliers and noise. During hyperparameter optimization, ‘class_weight="balanced"’ was set to handle sample imbalance. The screening data was partitioned into a training set of 2928 cases (including the validation set) and a test set of 325 cases at an 8:1:1 ratio. Python was used to facilitate data transformation. Eight different ML algorithms—Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), Gradient Boosting (GBM), Extreme Gradient Boosting (XGB), Deep Neural Network (DNN), Decision Tree (DT) and Naive Bayes (NB)—were used to construct a threatened abortion prediction model. The prediction performances of the ML models were evaluated by calculating the area under the curve (AUC) values. We used the SHapley Additive exPlanation (SHAP) method to explain the models.ResultsComparatively, the DNN model showed the highest predictive performance among the eight models, with the highest AUC value of 96.76% and top metrics for accuracy (91.88%), specificity (91.62%), sensitivity (92.11%), and F1 score (92.48%). SHAP analysis identified Red Cell Distribution Width - Standard Deviation (RDW-SD), Platelet Distribution Width (PDW), Mean Platelet Volume (MPV), Red Cell Distribution Width - Coefficient of Variation (RDW-CV), Absolute Basophil Count (BAS#), Platelet Count (PLT), Mean Corpuscular Hemoglobin Concentration (MCHC) and Lymphocyte Percentage (LYM) as the most influential features in predicting threatened abortion, with PDW, RDW-CV, BAS#, PLT, MCHC and LYM positively contributing to the prediction, whereas RDW-SD and MPV had negative contributions.ConclusionsOur research on constructing a prediction model for threatened abortion through routine blood tests has revealed the great potential of ML algorithms in detecting threatened abortion. This algorithm is expected to analyse routine blood data to identify at-risk pregnancies at an early stage, significantly improving the early detection of this common pregnancy complication. It will assist healthcare providers in intervening earlier and reducing the incidence of abortion. However, before the model can be translated into routine clinical applications, more extensive validation studies are still needed.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12884-025-08030-z.

ObjectivesNearly 20% of pregnancies in patients with SLE result in an adverse pregnancy outcome (APO). We previously developed an APO prediction model using logistic regression and data from Predictors of...

Septic shock is a severe complication of acute pancreatitis (AP), often associated with poor prognosis. This study aims to analyze the clinical characteristics of patients with acute pancreatitis and develop an interpretable early prediction model for septic shock in these patients using machine learning (ML). The model is intended to assist emergency physicians in resource allocation and medical decision making. Data were collected from the MIMIC-IV 3.0 database. The dataset was divided into a training set and a test set in a 7:3 ratio. Feature selection was performed using LASSO (Least Absolute Shrinkage and Selection Operator) regression. Subsequently, 10 ML models were developed: Random Forest, Logistic Regression, Gradient Boosting Machine, Neural Network, Extreme Gradient Boosting (XGBoost), K-Nearest Neighbor, Adaptive Boosting, Light Gradient Boosting Machine, Category Boosting, and Support Vector Machine. To enhance and optimize model interpretability, Shapley Additive Explanations (SHAP) were employed. A total of 1032 patients with AP were included in this study, from which 31 variables were selected for model development. By comparing the area under the receiver operating characteristic curve and decision curve analysis results between the training and test sets, the XGBoost model demonstrated a significant advantage over other models. SHAP analysis revealed that white blood cell count, total bilirubin (bilirubin total), and bicarbonate (HCO3 -) levels were the three most critical risk factors for the development of septic shock in patients with AP. ML approaches exhibited promising performance in predicting septic shock in patients with AP. These models may aid in guiding treatment decisions for patients with AP in the emergency department.

ObjectiveAnalyze the psychological and clinical factors of clinically significant tinnitus (THI score ≥38) in patients with hearing loss, construct predictive models based on four machine learning (ML) algorithms, and compare the predictive performance of different models.MethodsPatients with hearing loss who visited the Department of Otolaryngology at Qinghai University between August 2024 and May 2025 were enrolled in this study. Clinical data were retrieved from the hospital’s electronic medical record system. The study outcome was the occurrence of clinically significant tinnitus. Predictive variables were screened using univariate analysis, the least absolute shrinkage and selection operator (LASSO) regression, and the Boruta algorithm. Four ML algorithms—logistic regression (LR), random forest (RF), extreme gradient boosting (XGBoost), and support vector machine (SVM)—were applied to construct and validate predictive models. The area under the receiver operating characteristic curve (AUC) of each model in the validation set was compared using the DeLong test. Additionally, model performance metrics in the validation set were compared to identify the optimal model. Finally, the Shapley additive explanations (SHAP) algorithm was employed to interpret the best-performing model.ResultsNine key variables—age, hypertension, sleep disorder, anxiety, hearing loss severity, depression, noise exposure history, hearing side, and ototoxic drug use—were retained after LASSO and Boruta feature selection. Among the four ML models, the RF algorithm achieved the best predictive performance, with an AUC of 0.973 in the training set and 0.977 in the validation set, followed by XGBoost (AUC = 0.962 and 0.961, respectively). DeLong tests confirmed that RF significantly outperformed LR and SVM models (p < 0.001), while its difference from XGBoost was not significant. In the validation set, the RF model yielded the highest accuracy (0.923), sensitivity (0.929), specificity (0.914), precision (0.945), and F1-score (0.937). SHAP analysis indicated that hearing loss severity, age, and sleep disorder were the most influential predictors, suggesting that both auditory and non-auditory factors contribute substantially to the risk of clinically significant tinnitus.ConclusionThe RF model showed the best performance in predicting clinically significant tinnitus, with hearing loss severity, age, and sleep disorder identified as major predictors. Integrating auditory and psychological factors can improve early risk identification in patients with hearing loss.

Simple SummaryThe objective response rate of lenvatinib combined with transarterial chemoembolization for unresectable hepatocellular carcinoma is unsatisfactory. We aimed to develop predictive models using demographic characteristics, pre-treatment serum biomarkers and tumor characteristics of unresectable hepatocellular carcinoma patients by five machine learning algorithms to predict the response under combined treatments. We identified the 10 most important predictors, including K, low-density lipoprotein, D-dimer, red blood cell, alanine aminotransferase, albumin, monocyte, tumor size, triglyceride, and age. In addition, we applied the Shapley Additive exPlanation to explain the best-performing random forest predictive model to provide a reasonable explanation of the efficacy prediction at an individualized level. The combination of machine learning and Shapley Additive exPlanation can provide valuable suggestions for clinical decision making.Background: Lenvatinib and transarterial chemoembolization (TACE) are first-line treatments for unresectable hepatocellular carcinoma (HCC), but the objective response rate (ORR) is not satisfactory. We aimed to predict the response to lenvatinib combined with TACE before treatment for unresectable HCC using machine learning (ML) algorithms based on clinical data. Methods: Patients with unresectable HCC receiving the combination therapy of lenvatinib combined with TACE from two medical centers were retrospectively collected from January 2020 to December 2021. The response to the combination therapy was evaluated over the following 4–12 weeks. Five types of ML algorithms were applied to develop the predictive models, including classification and regression tree (CART), adaptive boosting (AdaBoost), extreme gradient boosting (XGBoost), random forest (RF), and support vector machine (SVM). The performance of the models was assessed by the receiver operating characteristic (ROC) curve and area under the receiver operating characteristic curve (AUC). The Shapley Additive exPlanation (SHAP) method was applied to explain the model. Results: A total of 125 unresectable HCC patients were included in the analysis after the inclusion and exclusion criteria, among which 42 (33.6%) patients showed progression disease (PD), 49 (39.2%) showed stable disease (SD), and 34 (27.2%) achieved partial response (PR). The nonresponse group (PD + SD) included 91 patients, while the response group (PR) included 34 patients. The top 40 most important features from all 64 clinical features were selected using the recursive feature elimination (RFE) algorithm to develop the predictive models. The predictive power was satisfactory, with AUCs of 0.74 to 0.91. The SVM model and RF model showed the highest accuracy (86.5%), and the RF model showed the largest AUC (0.91, 95% confidence interval (CI): 0.61–0.95). The SHAP summary plot and decision plot illustrated the impact of the top 40 features on the efficacy of the combination therapy, and the SHAP force plot successfully predicted the efficacy at the individualized level. Conclusions: A new predictive model based on clinical data was developed using ML algorithms, which showed favorable performance in predicting the response to lenvatinib combined with TACE for unresectable HCC. Combining ML with SHAP could provide an explicit explanation of the efficacy prediction.

Objective:The aim of this study was to develop and validate a machine learning model to predict the risk of comorbid depression in asthma patients.Methods:We conducted a retrospective study of 2464 asthma patients with comorbid depression using National Health and Nutrition Examination Survey (NHANES) data. Feature selection was conducted using the Boruta algorithm and the Least Absolute Shrinkage and Selection Operator (LASSO). Eight machine learning algorithms, namely Decision Tree (DT), k-Nearest Neighbors (KNN), Light Gradient Booster Machine (LGBM), Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), eXtreme Gradient Boosting (XGBoost), and Multilayer Perceptron (MLP), were trained using 5-fold cross-validation methodology. Model performance was evaluated through various metrics such as area under the curve (AUC), accuracy, sensitivity, specificity, F1 score, and decision curve analysis (DCA). Interpretation was conducted using SHapley Additive exPlanations (SHAP) analysis, highlighting feature importance.Results:The training set comprised 1724 participants, while the validation set included 740 participants, with a depression prevalence of 14.45%. Significant predictors identified included hypertension, chronic obstructive pulmonary disease (COPD), stroke, sleep questionnaire (SLQ) scores, smoking status, Poverty Index Ratio (PIR), and educational level. The XGBoost model demonstrated superior performance compared with alternative machine learning (ML) algorithms, achieving an AUC of 0.750, an accuracy of 69.1%, a sensitivity of 68.2%, a specificity of 73.8%, and an F1 score of 79%. The SHAP method identified SLQ, PIR, and education level as the primary decision factors influencing the ML model’s predictions.Conclusion:The XGBoost model effectively predicts the risk of depression in asthma patients, serving as a valuable reference for early clinical identification and intervention.

The study aimed to develop and externally validate multiparametric MRI (mpMRI) radiomics-based interpretable machine learning (ML) model for preoperative differentiating between benign and malignant prostate masses. Patients who underwent mpMRI with suspected malignant prostate masses were retrospectively recruited from two independent hospitals between May 2016 and May 2023. The prostate mass regions in T2-weighted imaging (T2WI) and diffusion-weighted imaging (DWI) MRI images were segmented by ITK-SNAP. PyRadiomics was utilized to extract radiomic features. Inter- and intraobserver correlation analysis, t-test, Spearman correlation analysis, and the least absolute shrinkage and selection operator (LASSO) algorithm with a five-fold cross-validation were applied for feature selection. Five ML learning models were built using the chosen features. Model performance was evaluated with internal and external validation, using area under the curve (AUC), calibration curves, and decision curve analysis to select the optimal model. The interpretability of the most robust model was conducted via SHapley Additive exPlanation (SHAP). A total of 567 patients were enrolled, consisting of the training (n = 352), internal test (n = 152), and external test (n = 63) sets. In total, 2,632 radiomic features were extracted from regions of interest (ROIs) of T2WI and DWI images, which were reduced to 18 via LASSO. Five ML models were established, among which the random forest (RF) model presented the best predictive ability, with AUCs of 0.929 (95% confidential interval [CI]: 0.885-0.963) and 0.852 (95% CI: 0.758-0.934) in the internal and external test sets, respectively. The calibration and decision curve analyses confirmed the excellent clinical usefulness of the RF model. Besides, the contributing relations of the radiomic features were uncovered using SHAP. Radiomic features from mpMRI combined with machine learning facilitate accurate preoperative evaluation of the malignancy in prostate masses. SHAP can disclose the underlying prediction process of the ML model, which may promote its clinical applications.