A retrospective prognostic evaluation using unsupervised learning in the treatment of COVID-19 patients with hypertension treated with ACEI/ARB drugs

Data collection and feature extraction

This study was approved by the Ethics Committee of Shanghai City Jiading District Hospital (2022-K10). Informed consent was obtained from all individual participants included in the study. This was a single-center, retrospective study of patients with COVID-19 from Tongjia Designated Hospital in Jiading District, Shanghai city. With the approval of the Ethics Committee of Shanghai City Jiading District Central Hospital, informed consent was waived. Patients with COVID-19 and hypertension were selected for the study. The inclusion period ranged from April 7, 2022 to May 19, 2022. A total of 200 patients (96 males and 104 females) were included. The final follow-up date was June 21, 2021.

The inclusion criteria for patients were as follows: (1) refer to the “Diagnosis and Treatment Plan for Novel Coronavirus Pneumonia (Trial Version 10)” (Nicola et al., 2020) and “Guidelines for the Prevention and Treatment of Hypertension in China (2018 Edition)” (Théry et al., 2018). According to the relevant diagnostic criteria in the study, (2) a comprehensive evaluation of the epidemiological history, clinical symptoms, laboratory indicators and imaging findings of the patients was performed; the exclusion criteria were as follows: (1) aged less than 18 years; (2) had SARS-CoV-2 asymptomatic infection; (3) had advanced malignant tumors and needed radiotherapy and chemotherapy; (4) had severe organ failure at admission; and (5) had complete organ failure.

We collected demographic information from two groups of patients, namely, clinical symptoms, diagnosis and classification, auxiliary inspection and clinical outcome.

Improved algorithm and synchronous optimization

The swarm intelligence optimization algorithm can provide a good foundation for feature selection and clustering analysis (Hanrahan, 2011; Li et al., 2022). The principle of the improved wild dog optimization algorithm (IDOA) is briefly introduced as follows. The name of our improved algorithm is IDOA-K-Means. First, the IDOA uses the cubic chaotic mapping strategy to initialize the population, and the chaotic mapping can be expressed as (1)${\displaystyle x_{n+1}=\rho x_n\left(1-x_n^2\right).}$

Here, x_n corresponds to the numerical value of the nth dimension. To enhance the traversal, we set the initial value to x₀ =0.3 and the parameter ρ =2.593. In chaotic mapping, the selection of parameter values needs to be determined based on specific requirements and system characteristics. For cubic chaotic maps, the parameter ρ affects the stability and periodicity of the system. When the parameter ρ is 2.593, the cubic chaotic mapping has the best traversal performance, which means that the system can better explore various points in the state space. The purpose of setting the initial value to 0.3 is to ensure that the values generated by cubic chaotic mapping fall between [0,1]. This initial value selection takes into account the randomness of the initial state and the rationality of the mapping results, allowing the algorithm to better explore potential solution spaces. Through such an initialization method, the blindness of population initialization can be overcome so that the population can better cover the search space.

Second, the IDOA uses dimensionwise Gaussian variation for the optimal individual. Specifically, for each dimension j, the optimal position is changed by Gaussian mutation, which can be expressed as (2)${\displaystyle Xbestnew\left(j\right)=w\ast Xbest\left(j\right)+randn\ast Xbest\left(j\right).}$

In the process of dimensionwise Gaussian mutation, an inertia weight w is used to balance the influence of the new solution and the old solution. Xbest represents the best position before updating, and Xbestnew represents the new position after Gaussian mutation.

Finally, the IDOA uses a greedy strategy to update the fitness; that is, it uses a greedy strategy to retain the optimal solution, which can be expressed as (3)${\displaystyle Xbest=\left\{\begin{array}{c}Xbestnew,iff\left(Xbestnew\right)<f\left(Xbest\right)\hfill \\ Xbest,else\hfill \end{array}.\right.}$

This strategy aims to improve the ability of the optimal individual to jump out of the local optimal solution, thereby enhancing the global search ability of the algorithm. This means that if the fitness of the new solution is better, update the optimal solution to the new solution; otherwise, keep the original solution unchanged. Through such a strategy, the best solution can be obtained, and the performance of the algorithm can be improved.

Hence, by introducing dimensionwise Gaussian mutation and a greedy strategy, the IDOA is beneficial for overcoming the local optimal problem of traditional optimization algorithms and improving the global search ability of the algorithm. The core idea of the IDOA is to increase the diversity of the population through the mutation strategy to better search for the optimal solution. Using the 23 common test function to evaluate the performance of the IDOA, as shown in Fig. 2, the results reveal that the IDOA is superior to the control method in terms of convergence speed and the ability to obtain the global optimal solution.

Feature selection and cluster analysis

After successfully constructing the IDOA, we use the global optimization ability of the IDOA to complete the two key steps of feature selection and optimal clustering center point optimization at the same time.

The Euclidean distance is used to calculate the clustering between the sample points and the cluster center; that is, the sum of squared errors (SSE) is used as the objective function for optimization, and the formula can be expressed as follows: (4)${\displaystyle SSE=\sum _{i=1}^k\sum _{X_m\in C_i}{∥X_m-C_i∥}^2.}$

Here, C_i represents the cluster center point, and X_m represents each sample data point. The cluster center point is updated through each iteration, and the update rule calculates the new center point coordinates according to the category to which the sample points belong. The smaller the SSE is, the better the clustering effect, and the clustering center point is updated through each iteration C_i. The update rule can be expressed as follows: (5)${\displaystyle C_i=\frac{1}{n}\sum _{m=1}^{n_i}{X}_m,i=1,2\dots ,k.}$

Here, n_i represents the number of data points in category i, and the cluster center point is updated through each iteration of the above formula until the cluster center does not change, indicating that the model converges effectively.

A flow chart of the synchronous optimization process is shown in Fig. 3. The initialization process determines the feature screening and the encoding method of the center point; the result of feature screening determines the dimension of the cluster center point; finally, the selected features and center point are merged, and optimization is performed on this basis. The detailed steps can be described as follows. First, in the initialization process, the feature screening part adopts 0–1 encoding, indicating whether the feature is retained, and the center point adopts real number encoding, indicating the coordinates of the center point. Second, 0 means to remove features, and 1 means to keep features. The dimensions of the center point are determined according to the number of features chosen. For example, three features are reserved in the above figure, so the first three real numbers are selected as the coordinates of the center point. Finally, the selected features and center point coordinates are merged, and the objective function is added to the merged dataset for optimization. Throughout the process, feature selection and central point optimization are carried out simultaneously. By merging the datasets and inputting them into the objective function, the optimal features and the best clustering central points can be obtained.

Model training and evaluation

To avoid the risk of overfitting and improve the stability of training, 50-fold cross-validation was applied to the training set. Subsequently, we evaluated the performance of the final model by selecting 80% of the total number of patients as the training set and the remaining 20% as the test set. This method can effectively avoid overfitting of the model on the training set and obtain more accurate prediction results on the test set.

The data were analysed using IBM SPSS (26.0) with a significance level set at 5%. Count data were compared using the t test (Liu & Wang, 2021), while categorical variables were compared using the χ² test (Aslam, 2021).

Feature selection and cluster analysis results

This section compares the traditional k-means algorithm with the IDOA-K-means algorithm and provides a two-dimensional scatter plot and a comparison of the prognostic performance of k-means clustering analysis to demonstrate the differences before and after algorithm optimization.

First, in the cluster analysis, we screened out six characteristics that affected the differences between groups, age, hypertension grade, MuLBSTA (Iijima et al., 2021a), use of ACEI/ARB, NT-proBNP, and high-sensitivity troponin I. The MuLBSTA score (Iijima et al., 2021b) is a valuable and effective early warning model that has been developed to predict the mortality rate of patients with viral pneumonia. These features included multiple infiltrations, hypophocytosis, bacterial infection, smoking history, hyper tension, and age. Afterwards, to better visualize the clustering results, we applied principal component analysis (PCA) to reduce the dimensionality of these 6 features. In this dimensionality-reduced feature space, we draw a two-dimensional scatter plot, as shown in Fig. 4. The results of the scatter plot revealed that the clustering effect was good, and there was clear differentiation between the two groups of patients with good and poor prognoses. This shows that the application of the IDOA-k-means algorithm to this dataset is successful and that the algorithm can be used to classify patients effectively and identify distinguishing features.

However, we also conducted a prognostic performance analysis on the two groups processed by clustering analysis. The results before optimization are shown in Table 1, and the results after optimization are shown in Table 2. The results before improvement showed that k-means clustered the data into two categories, but there was no significant difference in the prognosis between the two groups of patients, which was not sufficient. The improvement results showed that the hospitalization days and nucleic acid conversion time in group 2 were greater than those in group 1, while the nucleic acid positivity rate (27.03%) in group 2 was greater than that in group 1. The differences in these indicators were statistically significant (P < 0.05). In addition, we also observed that there were no deaths in group 1, while there were 2 deaths in group 2. Based on these observations, we classified group 1 as the good prognosis group and group 2 as the poor prognosis group.

Difference analysis of key characteristics between the two groups of patients

This section analyses the differences in key characteristics of the two groups of patients from the aspects of baseline information, clinical symptoms, laboratory tests, and imaging examinations and then selects the key characteristics that affect the differences between the two groups, as shown in Supplementary Material S1.

First, for the baseline data, we observed that there were significant differences between the two groups in age, ACEI/ARB drug use, hypertension grade, new crown type and MuLBSTA (P < 0.05). These features are closely related to patient prognosis. Moreover, the clinical symptom indices exhibited minor discrepancies between the two patient groups and did not demonstrate statistical significance (P > 0.05). Hence, clinical symptom indicators were not the key factors in predicting the prognosis of patients in this study. Third, regarding the laboratory test indicators C-reactive protein (CRP), interleukin-6 (IL-6), lactate dehydrogenase, D-2 polymer, oxygenation index, procalcitonin (PCT), aspartate aminotransferase (AST), creatinine, NT-proBNP, and hypersensitivity, there were statistically significant differences in lymphocyte counts between patients in the two groups. Additionally, there were significant differences in troponin I and cycle threshold of COVID-19 nucleic acid detection (P < 0.05). These laboratory test results may be highly important for predicting patient prognosis. Finally, in terms of imaging examination indices, there was no significant difference in the results of X-ray examination between the two groups (P > 0.05), which suggested that X-ray examination is useful for evaluating patient prognosis.

Ranking and verification of feature importance

According to the results of the 17 features with statistical significance in the feature difference analysis, we used the ReliefF algorithm (Zhang, Ding & Li, 2008; Liu, Chen & Huang, 2023) to further rank the importance of these features, as shown in Fig. 5. In the ranking of the importance of features, the top six features were age, hypertension grade, MuLBSTA, use of ACEI/ARB, NT-proBNP, and high-sensitivity troponin I. The importance ranking results of these features are consistent with the results of unsupervised prediction model screening features, as demonstrated in the first part of the Results section (“Feature selection and cluster analysis results”). This further verified the correlation of these features with the prognosis of patients with new crowns. The ranking of the importance of these features can provide a reference for clinicians to help them more accurately assess the prognosis of patients and take corresponding treatment measures.

To verify the importance of these selected features, we combined the six features screened out by the model into a multifactor logistic regression model, as shown in Table 3. To address categorical variables, we turned them into dummy variables. The outcome variable was defined as the occurrence or absence of an adverse clinical outcome, where 0 indicates a good prognosis and 1 indicates a poor prognosis. After stepwise screening, we found that age (OR = 1.055, P = 0.009) and ACEI use (OR = 0.302, P = 0.020) were two significant independent variables.

Visualization system construction

The aforementioned studies have successfully identified key features that influence patient outcomes. However, in clinical practice, the changes in these characteristics are intricate, and it is difficult to intuitively determine the prognostic risk. Existing artificial intelligence methods have high barriers to popularization and application, requiring clinicians to have high programming skills and extensive literature knowledge, which makes them difficult to promote and use in a large number of hospitals. To solve this problem, this article innovatively constructs a practical visualization system that is built on the basis of selected key features and has the advantages of intuition, convenience and practicability. During the application of the visualization system, the user needs to input only the specific values of the six key features in the “baseline information” column, and the system can automatically calculate the prognosis level and provide targeted suggestions. Figure 6 shows an example of an application for prognostic assessment. In the process of clinical diagnosis and treatment, this system helps to quickly screen for prognostic risk and take timely targeted measures to address high-risk patients. The construction of this system helps patients achieve timely clinical intervention, thereby reducing the risk of adverse outcomes, which has important practical significance and application value.

The code, software version, and parameters for “feature selection and clustering analysis” are available in GitHub: https://github.com/gelihua/IDOAKmeans. The difference analysis of key characteristics between two groups of patients was conducted using SPSS 26.0. Use MATLAB R2022a’s built-in Train Classification Models in Classification Learner App for visualization of feature importance sorting and validation. The functional parameter selection is FeatureSelection ReliefF.

The code, software version, and parameters of the “visualization system” can be found on GitHub: https://github.com/gelihua/IDOAKmeans. This software is developed based on MATLAB R2022a and utilizes the APP Designer function to design the original *. mlapp file. On this basis, MATLAB’s built-in MATLAB Runtime compiler is used to compile the *. mlapp file into a *. exe file that can be executed independently of the MATLAB environment. As long as the software is pre installed on a computer with MATLAB Runtime, it can be run, successfully reducing the running environment requirements of the software and improving portability. (1) Hardware requirements: Processor main frequency: 2Ghz and above; memory: 2GB or above. (2) Software requirements. System: Windows 10 and above (64-bit system); operating environment: MATLAB R2022a and above. MATLAB Runtime version 9.12 and above.