A prediction model of outcome of SARS-CoV-2 pneumonia based on laboratory findings

Abstract

The severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) has resulted in thousands of deaths in the world. Information about prediction model of prognosis of SARS-CoV-2 infection is scarce. We used machine learning for processing laboratory findings of 110 patients with SARS-CoV-2 pneumonia (including 51 non-survivors and 59 discharged patients). The maximum relevance minimum redundancy (mRMR) algorithm and the least absolute shrinkage and selection operator logistic regression model were used for selection of laboratory features. Seven laboratory features selected in the model were: prothrombin activity, urea, white blood cell, interleukin-2 receptor, indirect bilirubin, myoglobin, and fibrinogen degradation products. The signature constructed using the seven features had 98% [93%, 100%] sensitivity and 91% [84%, 99%] specificity in predicting outcome of SARS-CoV-2 pneumonia. Thus it is feasible to establish an accurate prediction model of outcome of SARS-CoV-2 pneumonia based on laboratory findings.

Artificial intelligence in clinical care amidst COVID-19 pandemic: A systematic review

“…of CPP* AI methods Predictors Val. methods Performance (AUC, Accuracy (Acc%), Sensitivity (SEN%), Specificity (SPE%), PPV/NPV (%), (95% CI)) Risk of Bias**: Participants/Predictors/Outcome/Analysis/Overall Cai et al [83] , China, Severity assessment, ICU need and length of stay prediction, O 2 inhalation duration prediction, sputum NAT-positive prediction and patient prognosis 99 ML, RF CT quantification 10-FCV AUC 0.945 (ICU treatment), AUC 0.960 (prognosis/ partial recovery vs pro-longed recovery) H U U H H Wu et al [84] , China, Mortality prediction 58 ML, LR, mRMR, LASSO LR 7 continuous laboratory variables: blood routine test, serum biochemical (including glucose, renal and liver function, creatine kinase, lactate dehydrogenase, and electrolytes), coagulation profile, cytokine test, markers of myocardial injury, infection-related makers, other enzymes 5-FCV SEN 98 (95% CI: 93–100), SPE 91 (95% CI: 84–99) H L L H H Iwendi et al [27] , Severity and outcome prediction unclear ML, DT, SVM, GNB, Boosted RF, AdaBoost COVID-19 patient's geographical, travel, health, and demographic data unspecified Acc 94 H U H H H Gerevini et al [15] , Italy, Mortality prediction unclear DT, RF, ET (extra trees) Age, sex, C-Reactive Protein (PCR), Lactate dehydrogenase (LDH), Ferritin (Male)Ferritin (Female), Troponin-T, White blood cell (WBC), D-dimer, Fibrinogen, Lymphocyte (over 18 years old patients), Neutrophils/Lymphocytes, Chest XRay-Score (RX) Cross validation AUC 90.2 for the 10th day H U U L H Yan et al [108] , China, Severity assessment 404 ML, XGBoost 3 biomarkers that predict the survival of individual patients: LDH, lymphocyte, high-sensitivity C-reactive protein (hs-CRP). cross-validation Acc 90 L L ...…”

“…In particular, AI techniques, including ARMED (Attribute Reduction with Multi-objective Decomposition Ensemble optimizer) [62] , GFS (Gradient boosted feature selection) [62] , MRMR (Maximum Relevance Minimum Redundancy) [84] , and RFE (Recursive Feature Elimination) [62] were used for feature selection. Regression models, including L1LR (L1 Regularized Logistic Regression) [17] , LASSO (Least Absolute Shrinkage and Selection Operator) [50] , [59] , [84] , [89] , [112] , CoxPH (Cox Proportional Hazards) [100] , and SRLSR (Sparse Rescaled Linear Square Regression) [62] were also employed for feature selection. The use of Nadam optimizer (NesterovAccelerated Adaptive Moment optimizer) for model optimization was reported in two studies [33] , [117] , while application of SMOTE (Synthetic Minority Oversampling TEchnique) for data augmentation was reported in one study [58] .…”

“…In the participants domain, 30 studies had high risk of bias and 21 unclear. Sources of bias in the participants domain varied from small or incomplete datasets to exclusion criteria indicating the need of further data collection to test the generalizability of the developed AI models to other patient populations [2] , [7] , [15] , [16] , [17] , [35] , [37] , [43] , [48] , [57] , [59] , [65] , [69] , [71] , [72] , [73] , [74] , [75] , [76] , [77] , [78] , [79] , [80] , [81] , [82] , [83] , [84] , [85] , [86] , [87] , [88] , [89] , [90] , [91] . Thirty studies had high RoB in the “predictors” domain related to different ways of definitions and assessment for all participants or predictors availability.…”

“…Limitations related to the applicability of the developed prediction models were reported by several studies. The most prominent limitation reported was the use of a single data source (one hospital from one geographical area) for the algorithm’s training [4] , [10] , [11] , [15] , [17] , [30] , [56] , [58] , [59] , [62] , [63] , [64] , [69] , [44] , [76] , [83] , [84] , [91] , [92] , [78] , [103] , [107] , [106] , [108] , [109] , [112] , [114] , [115] , [116] , [118] . Generalizability of the trained models can be enhanced by adding multiple data sources in future studies.…”

Artificial Intelligence in Surveillance, Diagnosis, Drug Discovery and Vaccine Development against COVID-19

“…The XGBoost algorithm has great interpretability potential due to its recursive tree-based decision system and is shown to be ~90% accurate in predicting patient mortality approximately 2 weeks in advance [ 124 ]. Similarly, in another study, SARS-CoV-2 induced pneumonia was predicted based on seven laboratory parameters (prothrombin activity, urea, white blood cell, interleukin-2 receptor, indirect bilirubin, myoglobin, and fibrinogen degradation products) [ 125 ]. These parameters were identified by applying the least absolute shrinkage and selection operator (LASSO) logistic regression model based on features selected by the mRMR algorithm.…”