PAWNN score predicts hospital death risk of new crown patients

Click [Medical side] Follow us today to share with you a new paper published by researchers from Wuhan University in Med magazine, a subsidiary of Cell on January 7, 2021.
The author uses machine learning algorithms to screen important blood routine indicators to establish a comprehensive The risk scoring model (PAWNN) is used to predict the death risk of COVID-19 hospitalized patients, which will help optimize clinical decision-making and rationally allocate tight medical resources.

This article has greatly inspired the scientific research of clinicians.
The data in this article is derived from conventional blood test indicators, combined with simple machine learning algorithms to develop an accurate prediction model.
It is a classic article about machine learning combined with routine clinical tests.
The papers on testing indicators are also worthy of in-depth study and excavation by everyone, hereby share with everyone, if there are any deficiencies, please criticize and correct.

Research background In 2021, the new coronavirus disease (COVID-19) is still raging around the world.
Neither developing or developed countries have escaped the clutches of the new coronavirus.
The prevention and control of COVID-19 has given the health system of countries with insufficient medical resources.
Caused significant pressure.

Therefore, there is an urgent need to develop an accurate and reliable risk assessment tool to assess the prognosis of the disease, and to help first-line clinicians optimize medical interventions and limited medical resources.

For this reason, the author collected and analyzed the complete blood count (CBC) longitudinal data of a large number of COVID-19 cases, and found that a comprehensive score based on some selected CBC parameters can dynamically predict the risk of impending death during hospitalization with high accuracy.

The author also revealed the longitudinal trajectory of CBC parameters and the comprehensive score of the severity of COVID-19 during hospitalization.

The results of its research will be particularly helpful in optimizing clinical decision-making and may reduce mortality in countries that are suffering from severe shortages of medical resources.

The data and methods totaled 12,911 COVID-19 patients, and 152 leukemia patients were excluded.

The remaining 12,759 patients were used for analysis, of which 9810 patients had undergone at least two CBC tests during the hospital stay and were designated as the training set, and one CBC test was designated as the validation set, and finally the generalized linear mixed model (GLMM) was used.
Choose risk factors and build models.

3174 patients who underwent at least 3 CBC tests at 3 different stages during hospitalization were included in the Latent Markov Model (LMM) study.

In addition, there are 227 queues from Italy that have the ability to verify the PAWNN score.

Result display 1.
Analysis of clinical characteristics of inpatients in Hubei Province Table 1 shows the baseline clinical characteristics, past chronic diseases and laboratory tests of the patients upon admission.

2.
Dynamic trajectory of CBC parameters Figure 2 shows the linear fitting curve of the dynamic trajectories of 13 CBC parameters grouped by disease severity from admission to the 30th day of hospitalization.

White blood cell (WBC) count, neutrophil count, neutrophil percentage, and neutrophil/lymphocyte ratio (NLR) all increased with the severity of the disease.

In contrast, the levels of lymphocyte count, lymphocyte percentage, monocyte count, eosinophil count, basophil count, and platelet count decrease with the severity of the disease.

There were no significant differences in red blood cell (RBC) count, hematocrit, and hemoglobin concentration at the time of admission.

3.
Predictor selection and score development training cohort 9810 patients were used for variable selection and risk score formulation.

The variable selection process is shown in Figure 3.
A total of 38 variables are used for subsequent variable selection, including 25 categorical variables and 13 continuous parameters.

Generalized linear mixed effects models (GLMMs) with fixed effects of each variable are sorted according to the Akaike Information Criteria (AIC).

In order to further select the most effective fixed-effect predictors, the author used the multivariate GLMMs of the positive stepwise method to select 4 blood routine parameters, and finally added the category of control age.
Based on this, the Cox proportional hazard regression model was used to form A risk assessment scoring model (PAWNN score) consisting of 5 variables.

The specific numerical scores of each factor can be seen in Table 2.

4.
The performance of PAWNN score in the training and validation cohort The author conducted internal verification of the accuracy and specificity of PAWNN score on patients through 10-fold cross-validation.

The area under the receiver operating characteristic (AUROC) curve ranges from 0.
92 (95% CI 0.
91-0.
93) to 0.
93 (95% CI 0.
92-0.
94).

The trend graph of the PAWNN score showed significant differences in the levels of non-serious survivors, severe survivors, and death groups during hospitalization.

By dividing the follow-up period into quartiles, we found that the PAWNN score is still very accurate in predicting mortality at different time intervals.

The lowest AUROC for 0-1 days after admission was 0.
89 (95% CI 0.
88-0.
90), and the cut-off value was 6; the highest AUROC at 8 days after admission was 0.
94 (95% CI 0.
93-0.
94/0.
95), and the cut-off value was 6 Points (Figure S1C and Table S3).

In the validation data set of 2949 cases in Hubei Province where only one CBC test was performed during hospitalization, the PAWNN score AUROC was 0.
97 (95% CI 0.
96-0.
98), the sensitivity was 93.
84% (95% CI 90.
51-98.
10), and the specificity was 90.
90 % (95%CI 85.
13-92.
84) (Table 3).

The performance of the PAWNN score was further tested in a cohort of COVID-19 patients from Milan, Italy, who collected CBC data on admission.

The predictability of the Italian cohort is still very high, with an AUROC of 0.
80 (95% CI, 0.
74-0.
86), a sensitivity of 68.
83% (95% CI, 58.
44-94.
81), and a specificity of 80.
67% (95% CI, 49.
33) -87.
33) (table 3).

5.
Latent Markov Model (LMM) Latent Markov Model (LMM) is composed of a structural model of a latent disease state and a measurement model of observation indicators.
These two models are the selected parameters in GLMM.

Figure 4 shows the structure of the three state models and the prevalence of each state.

The transition trajectory shows a clear pattern, that is, all patients with death outcomes either go through the process of intermediate to high risk, or directly return from the low-risk group, while a small number of intermediate-risk patients return to the low-risk state.

The transition probability from the low-risk state of Time1 to the medium-risk and high-risk states of Time2 is 18% and 1%, respectively.

The transition probability from the middle-risk group of Time1 to the low-risk group of Time2 is 27%.

In Time2, 17% of patients in the low-risk group and 1% of patients in the intermediate-risk group were transferred to the high-risk group of Time3.

In general, the PAWNN score of the high-risk group was significantly higher than that of the survival group at all time points.

PAWNN score is a good predictor of death at all three time points.
AUROC is 0.
77 at Time1, 0.
91 at Time2, and 0.
97 at Time3.

The AUROC of PAWNN score for the discrimination ability of Time2 latency and Time3 latency are 0.
86 and 0.
81, respectively.

For the predicted transition probability, the latency AUROC value from the PAWNN score of Time1 to Time2 is 0.
86, and the latency AUROC value from the PAWNN score of Time2 to Time3 is 0.
81.

Research conclusions The authors conducted a retrospective cohort study on a total of 13138 COVID-19 inpatients.
The generalized linear mixed model (GLMM) was used to select the complete blood count (CBC) as a potential predictor, and the Cox proportional hazard regression model was used to determine the platelet count.
A comprehensive score (PAWNN score) is obtained from five risk factors, age, white blood cell count, neutrophil count, and neutrophil/lymphocyte ratio.

The PAWNN score showed good accuracy in predicting mortality in the subgroups of 10-fold cross-validation (AUROCs 0.
92-0.
93) and different quartile follow-up intervals and previous diseases.

The effectiveness of the score was further verified in 2949 patients (AUROC 0.
97) and 227 patients in the Italian group (AUROC 0.
80).

The Hidden Markov Model (LMM) produces a recognizable patient state, where the PAWNN score is a distinguishing feature.

The PAWNN score has a good predictive ability for the transition probability between different potential conditions.

The PAWNN score is a simple and accurate risk assessment tool that can predict the mortality risk of COVID-19 patients during the entire hospital stay.

This tool can help clinicians prioritize the treatment of COVID-19 hospitalized patients, especially in underdeveloped areas with limited resources.

Reference: Liu, H.
, Chen, J.
, Development and Validation of a Risk Score Using Complete Blood Count to Predict In-hospital Mortality in COVID-19 Patients, Med (2021), doi: https://doi.
org /10.
1016/j.
medj.
2020.
12.
013.
END Call for Papers "Medical Formula" is now officially open to fans! The content must be originally published and related to scientific research.
Once adopted, a generous reward (300-2000 yuan) will be given.
Please stamp for details.

"Medical side" has always been committed to serving "medical people", pushing the most cutting-edge and most valuable original clinical and scientific research articles to clinicians and scientific researchers.

The medical department has launched "Laboratory Basics", "SCI Writing Skills", "Document Intensive Reading and Analysis", "Medical English Easy", "National Natural Science Foundation of China", "Clinical Data Mining", "Gene Data Mining", "R Language Tutorial", "Medical Statistics", "Minimally Invasive Animal Experiment Training" and many other special courses, if you need to know the detailed tweets of the courses, you can follow the "Medical Party" public account and click on the "excellent topics" to enter the Tencent classroom: https://medfun .
ke.
qq.
com NetEase Cloud Classroom: http://study.
163.
com/u/ykt1467466791112 Customer Service Tel: 15821255568 Customer Service WeChat: yixuefang1234