Anemia Classification with Clinical Feature Engineering and SHAP Interpretation

Tree-based machine learning (ML) algorithms, such as Extra Trees (ET), Random Forest (RF), Gradient Boosting Machine (GBM), and XGBoost (XGB) are among the most widely used in early drug discovery, given their versatility and performance. However, models based on these algorithms often suffer from misclassification and reduced interpretability issues, which limit their applicability in practice. To address these challenges, several approaches have been proposed, including the use of SHapley Additive Explanations (SHAP). While SHAP values are commonly used to elucidate the importance of features driving models' predictions, they can also be employed in strategies to improve their prediction performance. Building on these premises, we propose a novel approach that integrates SHAP and features value analyses to reduce misclassification in model predictions. Specifically, we benchmarked classifiers based on ET, RF, GBM, and XGB algorithms using data sets of compounds with known antiproliferative activity against three prostate cancer (PC) cell lines (i.e., PC3, LNCaP, and DU-145). The best-performing models, based on RDKit and ECFP4 descriptors with GBM and XGB algorithms, achieved MCC values above 0.58 and F1-score above 0.8 across all data sets, demonstrating satisfactory accuracy and precision. Analyses of SHAP values revealed that many misclassified compounds possess feature values that fall within the range typically associated with the opposite class. Based on these findings, we developed a misclassification-detection framework using four filtering rules, which we termed "RAW", SHAP, "RAW OR SHAP", and "RAW AND SHAP". These filtering rules successfully identified several potentially misclassified predictions, with the "RAW OR SHAP" rule retrieving up to 21%, 23%, and 63% of misclassified compounds in the PC3, DU-145, and LNCaP test sets, respectively. The developed flagging rules enable the systematic exclusion of likely misclassified compounds, even across progressively higher prediction confidence levels, thus providing a valuable approach to improve classifier performance in virtual screening applications.

The number of deaths from cardiovascular disease is projected to reach 23.3 million by 2030. As a contribution to preventing this phenomenon, this paper proposed a machine learning (ML) model to predict patients with arteriosclerotic heart disease (AHD). We also interpreted the prediction model results based on the ML approach and deployed modelagnostic ML methods to identify informative features and their interpretations. We used a hematology Electronic Health Record (EHR) with information on erythrocytes, hematocrit, hemoglobin, mean corpuscular hemoglobin, mean corpuscular hemoglobin concentration, leukocytes, thrombocytes, age, and sex. To detect and predict AHD, we explored random forest (RF), XGBoost, and AdaBoost models. We examined the prediction model results based on the confusion matrix and accuracy measures. We used the Shapley Additive exPlanations (SHAP) framework to interpret the ML model and quantify the contribution of features to predictions. Our study included data from 6,837 patients, with 4,702 records from patients diagnosed with AHD and 2,135 records from patients without an AHD diagnosis. AdaBoost outperformed RF and XGBoost, achieving an accuracy of 0.78, precision of 0.82, F1-score of 0.85, and recall of 0.88. According to the SHAP summary bar plot method, hemoglobin was the most important attribute for detecting and predicting AHD patients. The SHAP local interpretability bar plot revealed that hemoglobin and mean corpuscular hemoglobin concentration had positive impacts on AHD prediction based on a single observation. ML models based on real clinical data can be used to predict AHD.

Background: HIV/AIDS patients often experience hematological abnormalities like anemia, thrombocytopenia, neutropenia, and leukopenia, which require understanding for effective care and management.Therefore, the present research was designed to study the hematological profile of people living with HIV/AIDS (PLHA).Methods: A study involved 120 inpatient and outpatient patients who underwent comprehensive clinical, hematological, radiological, and microbiological investigations.Patients were given two blood samples for CD4 count analysis and hematological parameters assessment.Hematologic values were calculated based on specific cutoffs, including anemia (Hb<13% for men and Hb<12 gm% for non-pregnant women), thrombocytopenia (platelet count <1.5 lakh/mm 3 ), and leucopenia (white blood cell count <4000/mm 3 ).Mean corpuscular volume, hemoglobin, and concentration assessments were also conducted. Results:The study involved 120 patients, with 53.3% being male and 46.7% female.The mean corpuscular volume (MCV) was the highest, with a maximum of 38.3% in the 80-100 fml range, indicating anemia.The mean corpuscular hemoglobin (MCH) was the most common, with 55% having MCH between 27-32 picogm.The mean corpuscular haemoglobin concentration (MCHC) was the most common, with 52.5% having MCHC between 31-35%.The highest rectic count was 93.3%, with a TLC between 4000-11000/Cumm (94.2%).The study also showed that 49.2% of patients had normocytic normochromic anaemia, 17.5% had microcytic hypochromic anaemia, and 25% had macrocytic anaemia.Furthermore, various haematological parameters showed statistically significant associations with CD4 count levels among individuals studied.Conclusion: Hematological abnormalities, including anemia, are prevalent in patients with PLHA, contributing to their morbidity.Although ART may be responsible for anemia, the study found all patients on ART, making it difficult to prove this definitively.Identifying and correcting these abnormalities is crucial for improving patients' quality of life.

Objective: To explore the predictive values of routine blood test results for iron deficiency (ID) screening in children. Methods: Routine blood test results and serum ferritin (SF) levels from 1 443 healthy children (862 boys, 581 girls) aged 6 months to 18 years, who were seen for well-child visits between June 2017 and May 2019 in Children's Hospital, Zhejiang University School of Medicine, were retrospectively analyzed. ID was defined as SF<20 μg/L, iron deficiency anemia (IDA) as ID with anemia (hemoglobin(Hb)<110 g/L at 6 months-5 years of age, Hb<120 g/L at 6-18 years of age), non-anemia ID as ID without anemia, non-ID anemia as SF≥20 μg/L with anemia, and healthy control subjects as those with SF≥20 μg/L but without anemia. The blood test results including Hb, mean corpuscular volume (MCV), mean corpuscular hemoglobin concentration (MCHC), red blood cell distribution width (RDW), and the percentage of low hemoglobin density (LHD) of healthy control, non-anemia ID, non-ID anemia, and IDA groups were compared by analysis of variance (ANOVA) or non-parametric test, quantitative data were described as x±s or M(interquartile range), and receiver operating characteristic curve (ROC) analysis was applied to assess predictive values of routine blood test results and LHD for detecting IDA and ID. Results: Among 1 443 children with median age of 2.1(3.3) years, 1 061 children were in healthy control group, 292 in non-anemia ID group, 43 in non-ID anemia group and 47 in IDA group. The prevalence of ID was much higher than that of anemia (23.5% (339/1 443) vs. 6.2% (90/1 443) , χ(2)=169.76, P<0.01). Compared with control group, non-anemia ID group showed higher LHD (0.088 (0.093) vs.0.073 (0.068), P<0.01) and RDW (0.131±0.013 vs. 0.126±0.008, P<0.01), lower MCV ((80±4) vs. (83±4) fl, P<0.01) and MCHC values ((326±9) vs. (329±8) g/L, P<0.01). IDA group showed higher LHD (0.322(0.544)) and RDW (0.151±0.018), lower MCV ((73±6) fl) and MCHC values((309±14) g/L) than non-anemia ID group (all P<0.01). The area under curve (AUC) values of MCHC, LHD, RDW and MCV for detecting ID were 0.63 (95%CI: 0.60-0.67), 0.63 (95%CI:0.60-0.67), 0.67 (95%CI: 0.63-0.70) and 0.73 (95%CI: 0.69-0.76) respectively. With cutoff limits (MCV<80.2 fl, RDW>0.131 or MCHC<322 g/L), MCV, RDW and MCHC showed higher sensitivity for screening ID than hemoglobin (0.540, 0.469 and 0.336 vs. 0.139, χ(2)=121.70, 87.47, 35.56, all P<0.01). Conclusion: MCV, RDW and MCHC can be used to screen ID in primary health care settings.

An improved permeability estimation model using integrated approach of hybrid machine learning technique and Shapley additive explanation

Erythrocyte mean cell volume (MCV) is used clinically to classify anemia, and normal values may be used to exclude iron deficiency. We have studied the diagnostic accuracy of MCV and the related measures mean cell hemoglobin (MCH) and mean cell hemoglobin concentration (MCHC) in diagnosing empty iron stores in children and young adults. Diagnostic accuracy of MCV, MCH, and MCHC was studied by ROC curve analysis in 6443 ambulant patients aged 0.5-25 years, of which 476 were anemic. In all patients, blood hemoglobin, MCV, MCH, and serum ferritin were measured in specimens sampled at the same time. MCHC was calculated as MCH divided by MCV. The gold standard of empty iron stores was s-ferritin <10, 15, or 20 μg/L. The cutoff limit of MCV giving 90% sensitivity in diagnosing serum ferritin <15 μg/L was constructed using quantile regression. Generally, MCH was slightly more accurate than MCV and MCHC. In the whole study population, the area under the ROC curve was 0.68-0.93 for MCV, 0.73-0.96 for MCH, and 0.68-0.87 for MCHC; and 0.70-0.86, 0.71-0.89, and 0.68-0.88, respectively, in the anemic subpopulation. At the cutoff limits of MCV giving a sensitivity of 90% at all ages in anemic patients, the specificity was about 50%. Mean cell hemoglobin, MCH, and MCHC are only moderately accurate in diagnosing empty iron stores in children and young adults, and normal values of these tests do not exclude empty iron stores in anemic patients.

Machine learning technology that uses available clinical data to predict diabetic retinopathy (DR) can be highly valuable in medical settings where fundus cameras are not accessible. This study aimed to develop and compare machine learning algorithms for predicting DR without fundus image. We used data from Korea National Health and Nutrition Examination Survey (2008-2012 and 2017-2021) and enrolled individuals aged ≥ 20 years with diabetes who received fundus examination. Predictive models for DR were developed using logistic regression and three machine learning algorithms: extreme gradient boosting, decision tree, and random forest. Model performance was evaluated using area under the receiver operating characteristic curve (AUC) and accuracy for the diagnosis of DR, and feature importance was determined using Shapley Additive Explanations (SHAP). Among the 3,026 diabetic participants (male, 50.7%; mean age, 63.7 ± 10.5 years), 671 (22.2%) had DR. The random forest model, using 16 variables, achieved the highest AUC of 0.748 (95% confidence interval, 0.705-0.790) with a sensitivity 0.669, specificity of 0.729 and an accuracy of 0.715. As interpreted by SHAP, HbA1c, fasting glucose levels, duration of diabetes, and body mass index were identified as common key determinants influencing the model's outcomes. The DR prediction models using machine learning techniques demonstrated reliable performance even without fundus imaging, with the random forest model showing particularly strong results. These models could assist in managing DR by identifying high-risk patients, enabling timely ophthalmic referrals.

Introduction: Pulmonary artery (PA) pressure monitors have been shown to prevent hospitalizations in heart failure patients using mean pressure values. PA waveform analysis offers an overall assessment of right ventricular load that combines resistance and compliance. Hypothesis: We hypothesised that clinical outcomes can be predicted from PA waveform analyses from patients with an implanted PA monitor. Methods: The CHAMPION trial database was analysed. Information about the demographics and data collection methods for this cohort have been published previously. All waveforms were analysed automatically using Python. Tau or RC time, respiratory rate, respiratory range, P asymptote, mean pulmonary arterial pressure (mPAP), diastolic pulmonary arterial pressure (dPAP) and systolic pulmonary arterial pressure (sPAP) were derived from the raw waveforms. Tau and P asymptote were calculated using phase plane analysis. Outcomes of hospitalization or mortality were adjudicated as part of the original trial. Statistics were performed in R. A decision tree statistical approach using the XGBoost algorithm was used. Contribution to the model was assessed using Shapley Additive Explanations (SHAP) values. The dataset was split into 70% train and 30% test sets. Results: 296,948 PA pressure traces from 547 patients were analysed. 3 patients had insufficient quality of trace to analyse. There were 1340 outcome events throughout the follow up period, median 664 days. The trained algorithm was able to predict clinical outcomes in the test dataset with 100% accuracy. The most significant contributors according the mean SHAP value were RC time (SHAP = 1.1), respiratory range (SHAP = 1.1) and heart rate (SHAP = 1.1). Conclusions: Detailed analysis of the PA waveform from PA pressure monitors is feasible and yields clinically relevant information. A supervised machine learning algorithm trained on this data can accurately predict hospitalization or mortality events in heart failure patients.

Background: Acute kidney injury (AKI), with an increase in serum creatinine, is a common adverse drug event. Although various clinical studies have investigated whether a combination of two nephrotoxic drugs has an increased risk of AKI using traditional statistical models such as multivariable logistic regression (MLR), the evaluation metrics have not been evaluated despite the fact that traditional statistical models may over-fit the data. The aim of the present study was to detect drug-drug interactions with an increased risk of AKI by interpreting machine-learning models to avoid overfitting.Methods: We developed six machine-learning models trained using electronic medical records: MLR, logistic least absolute shrinkage and selection operator regression (LLR), random forest, extreme gradient boosting (XGB) tree, and two support vector machine models (kernel = linear function and radial basis function). In order to detect drug-drug interactions, the XGB and LLR models that showed good predictive performance were interpreted by SHapley Additive exPlanations (SHAP) and relative excess risk due to interaction (RERI), respectively.Results: Among approximately 2.5 million patients, 65,667 patients were extracted from the electronic medical records, and assigned to case (N = 5,319) and control (N = 60,348) groups. In the XGB model, a combination of loop diuretic and histamine H2 blocker [mean (|SHAP|) = 0.011] was identified as a relatively important risk factor for AKI. The combination of loop diuretic and H2 blocker showed a significant synergistic interaction on an additive scale (RERI 1.289, 95% confidence interval 0.226–5.591) also in the LLR model.Conclusion: The present population-based case-control study using interpretable machine-learning models suggested that although the relative importance of the individual and combined effects of loop diuretics and H2 blockers is lower than that of well-known risk factors such as older age and sex, concomitant use of a loop diuretic and histamine H2 blocker is associated with increased risk of AKI.

The high number of diabetes mellitus sufferers has become a global health issue, and a scientific approach is needed to produce accurate and efficient diagnoses, which can then support decision-making in providing solutions for its management. The goal of this research is to develop a machine learning model that can accurately, efficiently, and transparently diagnose diabetes mellitus for use in clinical practice. This research method involves using the Support Vector Machine (SVM) algorithm, optimized with the Grid Search technique, and evaluated interpretively using the SHapley Additive exPlanations (SHAP) method. This research uses a secondary dataset consisting of the parameters Pregnancies, Glucose, BloodPressure, SkinThickness, Insulin, Body Mass Index, DiabetesPedigree- Function, and Age. Data preprocessing was carried out by performing normalization using a standard scaler and dividing the data into training and testing sets. The results of this study show that the SVM model achieved an accuracy of 0.7532 with the optimal parameters C: 1, gamma: 0.01, and kernel: rbf. Using SHAP, the analysis shows that the parameters Glucose, Body Mass Index, and Age have a significant impact on the results of diabetes classification. The main finding of this study is that SupportVector Machine optimization with SHapley Additive exPlanations can deliver excellent performance in diabetes prediction while also enhancing model transparency. The study’s implications suggest that the results can serve as a foundation for developing a medical diagnosis system that is straightforward, accurate, and easy to understand for diabetes mellitus.

One hundred and forty-one blood samples of hemoglobin E (Hb E) carriers were collected to define suitable cutoff values of Hb E level, mean corpuscular volume (MCV), and mean corpuscular hemoglobin (MCH) as screening indicators for detecting individuals with double heterozygosity for Hb E/alpha(o)-thalassemia. Based on the values that gave 100% sensitivity, Hb E < 26%, MCV < 74 fl, and MCH < 24 pg were selected. Further validation of these selected values in additional 152 heterozygous Hb E pregnant women revealed 100% sensitivity, 86.2% specificity, and a 25.9% positive predictive value (PPV) for using Hb E cutoff point only, meanwhile, 100% sensitivity, 83.4% specificity, and 22.6% PPV were achieved for the MCV cutoff point. In addition, 100% sensitivity, 86.3% specificity, and 25.9% PPV were gained for the MCH cutoff point. Combining Hb E level with either MCV or MCH cutoff points, the specificity and PPV were increased to 95.1% and 50.0%, respectively. It is concluded that Hb E level < 26%, MCV < 74 fl, and MCH < 24 pg could be used for screening alpha(o)-thalassemia in heterozygous Hb E. However, to improve specificity and PPV of the tests, a combination of Hb E level < 26% with either MCV < 74 fl or MCH < 24 pg is recommended.

Elevated mean corpuscular volume (MCV) is common in persons with hemochromatosis associated with HFE C282Y homozygosity. We evaluated data from the subset of non-Hispanic white participants in the Hemochromatosis and Iron Overload Screening Study to determine if elevated MCV in C282Y homozygotes is related to this genotype or to serum iron measures. Regression analysis was used to model MCV and Hb from transferrin saturation (TfSat), serum ferritin (SF), mean corpuscular hemoglobin concentration, red blood cell count, age, HFE genotype, Field Center, and presence of liver-related abnormalities in C282Y homozygotes and control subjects without HFE mutations (wt/wt genotype). Mean MCV was higher in C282Y homozygotes than in HFE wt/wt controls (94.4 vs. 89.7 fL in women; 95.3 vs. 91.2 fL in men; P < 0.0001 for both). These differences were largely associated with increased mean TfSat and SF in C282Y homozygotes. Adjusted mean MCV was 92.0 fL (95% confidence interval, 91.1, 92.9) in female C282Y homozygotes and 90.9 fL (90.3, 91.5) in controls. Among women with SF in the reference range 20-200 microg/L, adjusted mean MCV was 92.9 fL, (91.7, 94.2) in C282Y homozygotes, 1.8 fL higher than in controls (P = 0.013). The adjusted mean MCV of male C282Y homozygotes and controls was similar (P = 0.30). Adjusted mean Hb was 0.2 g/dL higher in women with C282Y/C282Y than in controls. Greater mean MCV in C282Y homozygosity reflects increased mean TfSat and mean SF in men and women; an additional effect of genotype on MCV and Hb was detected in women.

Prediction of gully erosion susceptibility through the lens of the SHapley Additive exPlanations (SHAP) method using a stacking ensemble model.

Colorectal Cancer (CRC) is a highly prevalent malignancy globally, and early prediction is crucial for improving prognosis. This study used a multidimensional CRC dataset (n=1000) provided by the Kaggle platform, which contains 14 clinical and lifestyle characteristics. First, data imbalance was mitigated through Random Oversampling (ROM) and standardization. Subsequently, a comprehensive evaluation was performed on seven baseline machine learning models, including Gradient Boosting Decision Tree (GBDT), eXtreme Gradient Boosting (XGBoost) and so on. Based on performance metrics such as accuracy and F1 score, GBDT and XGBoost were subsequently selected as the optimal base learners. Finally, the predictive probability features generated by the base learners are fed into the meta-learners such as Random Forest (RF), K Nearest Neighbor (KNN) and Multi-Layer Perceptron (MLP) for secondary modeling. The interpretability of the model is achieved through the Shapley Additive exPlanations (SHAP) value, which quantifies the marginal contribution of each feature to the prediction. Experiments show that the RF integration architecture based on GBDT and XGBoost baseline models has the best performance (accuracy of 0.9527 and AUC of 0.9923). SHAP analysis showed that Activity_Level and BMI were core predictors with limited contribution from gender, confirming the prioritization of exercise and weight management in CRC prevention. The framework demonstrated excellent robustness and maintained its predictive advantage even when inefficient base models e.g., Logistic Regression (LR) were introduced. This study provides an interpretable machine learning paradigm for CRC risk stratification with potential for clinical translation.

Cardiovascular disease is the leading cause of mortality among nonalcoholic steatohepatitis (NASH) patients who undergo liver transplants. In the present study, machine learning algorithms were used to identify important risk factors for cardiovascular death and to develop a prediction model. The Standard Transplant Analysis and Research data were gathered from the Organ Procurement and Transplantation Network. After cleaning and preprocessing, the dataset comprised 10,871 patients and 92 features. Recursive feature elimination (RFE) and select from model (SFM) were applied to select relevant features from the dataset and avoid overfitting. Multiple machine learning algorithms, including logistic regression, random forest, decision tree, and XGBoost, were used with RFE and SFM. Additionally, prediction models were developed using a support vector machine, Gaussian naïve Bayes, K-nearest neighbors, random forest, and XGBoost algorithms. Finally, SHapley Additive exPlanations (SHAP) were used to increase interpretability. The findings showed that the best feature selection method was RFE with a random forest estimator, and the most critical features were recipient and donor blood type, body mass index, recipient and donor state of residence, serum creatinine, and year of transplantation. Furthermore, among all the outcomes, the XGBoost model had the highest performance, with an accuracy value of 0.6909 and an area under the curve value of 0.86. The findings also revealed a predictive relationship between features and cardiovascular death after liver transplant among NASH patients. These insights may assist clinical decision-makers in devising strategies to prevent cardiovascular complications in post-liver transplant NASH patients.