A population-based retrospective machine learning study of COVID-19 severity using integrated clinical and viral genomic data in Jiangsu Province, China

Abstract

Background

As coronavirus disease 2019 (COVID-19) has transitioned into an endemic phase characterized by sustained transmission and widespread hybrid immunity, understanding region-specific determinants of severe disease remains important for real-world risk stratification and public health planning.

Methods

A retrospective surveillance study was conducted using 5,072 severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) whole-genome sequences linked to clinical metadata from 13 cities in Jiangsu Province (from January 2023 to December 2024). Features derived from clinical, viral genomic, and regional epidemiological domains were evaluated using five machine learning models and assessed on an independent 2024 cohort. Model interpretability was examined using SHapley Additive exPlanations (SHAP) analysis. Key mutations were further examined through epitope prediction, peptide-HLA docking and binding affinity assessments to explore potential immunological implications.

Results

Integrated multidimensional features demonstrated superior predictive performance compared with single-domain inputs. In the independent 2024 validation cohort, LightGBM achieved the best overall performance (F1-score = 0.603; AUC = 0.735). SHAP analysis identified age as the dominant model predictor, followed by the age-viral load interaction, regional location, vaccination status, and selected viral genomic features. Epitope prediction and structural analyses suggested L452W-associated changes in predicted peptide-HLA interaction patterns within the evaluated set of high-frequency HLA class I alleles in the Jiangsu population, providing candidate hypotheses for future experimental validation.

Conclusions

COVID-19 severity during the endemic phase appeared to reflect interactions among host susceptibility, viral genetic variation, and regional epidemiological context, with age and vaccination emerging as key predictive factors. This population-based, interpretable framework highlights clinically relevant risk-associated features and may support real-world risk stratification in ongoing and future infectious disease surveillance.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12866-026-05360-6.

Background

With the continued evolution of severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) variants, COVID-19 has transitioned from a global pandemic to an endemic phase characterized by sustained local transmission [1]. Although widespread vaccination, hybrid immunity, and changes in variant pathogenicity have substantially reduced population-level mortality, sporadic severe cases continue to impose a persistent burden on healthcare systems [2]. Jiangsu Province, a densely populated economic and transportation hub in eastern China, provides a representative setting for investigating COVID-19 severity under endemic conditions. In the context of frequent variant turnover and ongoing community transmission, identifying key factors associated with severe disease remains essential for early risk stratification and optimized clinical intervention.

Previous studies have demonstrated that COVID-19 clinical outcomes are influenced by multiple factors, with viral genetic characteristics and host-related clinical features playing central roles [3, 4]. At the viral level, nonsynonymous mutations across the SARS-CoV-2 genome can modulate immune evasion, receptor-binding affinity, and tissue tropism, thereby potentially altering viral pathogenicity [5, 6]. At the host level, factors such as age, sex, vaccination history, and viral load have been consistently identified as major predictors of disease severity [4, 7]. However, marked variability in clinical outcomes persists even among individuals infected with the same viral variant, highlighting substantial inter-individual heterogeneity [8]. These observations suggest that variation in disease severity may be shaped by complex, potentially nonlinear interactions between viral genetic backgrounds and host-related factors, and such relationships may not be adequately captured by traditional statistical models when analyzing high-dimensional clinical and genomic data.

Recent advances in machine learning (ML) have provided powerful tools for integrating multidimensional data and identifying factors associated with severe COVID-19 outcomes. Several studies have demonstrated the utility of ML algorithms in clinical risk prediction. Yan et al. developed an interpretable XGBoost model using blood biomarkers to predict mortality [9], while Gao et al. constructed an ensemble-based early warning system achieving high accuracy [10]. To address class imbalance in severe outcome prediction, downsampling and ensemble strategies have proven effective [11]. These models can automatically learn nonlinear relationships and higher-order feature interactions, often outperforming traditional regression approaches. However, the “black-box” nature of many ML models raises concerns regarding clinical interpretability. The SHapley Additive exPlanations (SHAP) framework helps address this concern by decomposing complex predictions into additive feature contributions rooted in cooperative game theory, making model decisions more transparent [12].

Most existing studies, though, have drawn on global pandemic data with heterogeneous healthcare systems and immune backgrounds. While machine learning models based on clinical biomarkers or viral genomic information have demonstrated promising predictive performance, relatively few studies have jointly integrated clinical, viral genomic, and regional epidemiological information within a single interpretable framework. Population-specific analyses incorporating these complementary dimensions remain limited, particularly in East Asian populations during the endemic phase where virus-host interactions occur under widespread hybrid immunity.

In this study, we leveraged large-scale longitudinal surveillance data from Jiangsu Province, integrating SARS-CoV-2 whole-genome sequences, clinical metadata and regional epidemiological information collected from January 2023 to December 2024. By combining machine learning models with the SHAP interpretability framework, we aimed to: (1) systematically identify key factors associated with COVID-19 severity in an endemic setting; (2) characterize the relative contributions and interaction patterns of viral genetic features and host-related risk factors; and (3) assess model performance and interpret key predictive features through SHAP analysis. This work aims to provide data-driven insights to inform precision intervention strategies under endemic COVID-19 conditions and to demonstrate a generalizable analytical framework for assessing risk across viral variants. The overall pipeline of our COVID-19 severity prediction is summarized in Fig. 1.

Fig. 1.

Study design and analytical workflow

Materials and methods

Study design and data sources

We conducted a retrospective, population-based surveillance study using SARS-CoV-2 genomic and clinical metadata collected from 13 prefecture-level cities in Jiangsu Province, China, between January 2023 and December 2024. A subset of the 2023 sequences has been previously analyzed by our group to examine viral mutation patterns under host immune pressure [13]. The present study extends this work by incorporating 2024 surveillance samples and shifting the analytical focus to clinical severity prediction using machine learning approaches. Respiratory specimens were obtained through a standardized weekly surveillance program, including nasopharyngeal and oropharyngeal swabs. Each viral genome was linked to corresponding de-identified clinical metadata via a unique sample identifier. Clinical severity was classified according to the Protocol for Diagnosis and Treatment of COVID-19 (Trial 10th Edition) issued by the National Health Commission of the People’s Republic of China. Under this protocol, mild cases present with upper respiratory symptoms without radiological evidence of pneumonia, whereas moderate cases show pneumonic manifestations on chest imaging. Severe and critical cases are defined by respiratory or multi-organ compromise. For predictive modeling, disease severity was dichotomized into a non-severe group (asymptomatic and mild cases) and a severe-outcome group (moderate, severe, and critical cases). This grouping was based on clinical progression beyond mild infection, with moderate cases marking radiological pneumonia and lower respiratory tract involvement under the national protocol.

Metadata standardization and quality control

Clinical metadata from multiple surveillance sources were merged and standardized through a systematic preprocessing pipeline. Metadata variables were organized into four analytical domains: (i) demographic characteristics (age and sex); (ii) epidemiological factors (vaccination status and prior SARS-CoV-2 infection); (iii) spatiotemporal attributes (collection date and geographic regions); and (iv) clinical and virological parameters (clinical severity classification, viral variant type, and cycle threshold (Ct) values for the ORF1ab and N genes). Quality control procedures included: (i) standardization of age values by converting month-based entries to years where applicable; (ii) harmonization of categorical variables using uniform encoding schemas; (iii) aggregation of geographic regions into four economic zones following the commonly used regional classification adopted by Jiangsu Province based on geographic and socioeconomic characteristics [14]: southern Jiangsu (SuNan: Nanjing, Wuxi, Suzhou, Changzhou, and Zhenjiang), central Jiangsu (SuZhong: Nantong, Yangzhou, and Taizhou), northern Jiangsu (SuBei: Xuzhou, Lianyungang, Huai’an, Yancheng, and Suqian), and an “other” category including imported cases and records without city-level information; and (iv) removal of records with missing or invalid critical values, followed by deduplication based on unique sample identifiers.

Raw viral genomes were processed using Nextclade (v3.16.0) for lineage assignment and quality assessment. Genomic sequences were required to satisfy stringent criteria: (i) overall Nextclade QC status of “good”; (ii) genomic coverage ≥ 90%; (iii) ambiguous base content (N-content) ≤ 5%; (iv) sequence length within 95–105% of the SARS-CoV-2 reference genome; and (v) absence of internal premature stop codons. Qualified genomic data were then merged with the standardized clinical metadata via unique sample identifiers, yielding a high-quality integrated cohort for downstream multimodal analysis.

Feature engineering and selection

Genomic variations were encoded as binary amino acid substitution features in “Gene.Position” format, where each feature indicates the presence “1” or absence “0” of any amino acid substitution at that specific genomic position, regardless of the specific amino acid change. Viral clades were one-hot encoded. Per-gene mutation burden was computed by counting amino acid substitutions in each viral protein. To quantify structural impact, we calculated two physicochemical load metrics: (i) charge-change load, counting substitutions that altered amino acid electrostatic state (positive, negative, or neutral); and (ii) size-change load, counting substitutions that shifted side-chain van der Waals volume by at least one category (large: W, F, Y, R; medium: other residues; small: A, G, S) [15]. An interaction term (Age/Ct_N) was created to capture host-virus dynamics, reflecting the potential joint influence of host susceptibility and viral load on disease severity.

Following feature construction, categorical variables were one-hot encoded, and continuous variables were normalized via min-max scaling. The initial feature space consisted of 3,448 features, including 3,402 mutation features, 21 clade features, and 25 baseline clinical, regional or engineered features. To reduce dimensionality while retaining biologically meaningful signals, genomic features (mutation sites and clades) were subjected to a two-stage filtering procedure designed to improve temporal robustness. In the first stage, only genomic features with a variance of ≥ 1% were retained, reducing the mutation features from 3,402 to 28 and clade features from 21 to 1. In the second stage, the retained genomic features were filtered by their association with the binary severity outcome using a binary-feature association analysis based on the phi coefficient, which is equivalent to Pearson correlation for two binary variables (P < 0.05). This step retained 20 mutation features, while the remaining clade feature did not meet the significance threshold. Together with the 25 baseline clinical, regional, and engineered features that were retained without filtering, this yielded a final set of 45 features for subsequent model development.

Feature set construction and comparative analysis

To systematically evaluate the contribution of different feature categories, we organized the 45 features into three domains: Clinical (gender, age, Ct values, vaccination status, prior infection, Age_Ct_N interaction), Region (four economic zone indicators), and Mutation (selected mutation sites, per-gene mutation counts, physicochemical loads). Seven feature combinations (Clinical, Region, Mutation, Clinical+Region, Clinical+Mutation, Region+Mutation, and All) were constructed for comparative analysis.

The 2023 cohort (n = 2,240) was partitioned into training (75%, n = 1,680) and internal test (25%, n = 560) sets using stratified sampling. The 2024 cohort (n = 2,832) served as an independent temporal validation set. To address severe class imbalance, random downsampling was applied: training sets were balanced to 1:1 ratio, internal test sets to 1:1 ratio, and the 2024 validation set to 2:1 ratio.

Five machine learning algorithms (Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), XGBoost, LightGBM) were trained with default hyperparameters on balanced training sets and evaluated on both test sets. Performance was assessed using F1-score, area under the receiver operating characteristic curve (AUC), area under the precision–recall curve (AUPR), and accuracy. To identify the most robust feature combination independent of algorithm choice, we calculated Mean Reciprocal Rank (MRR) by averaging the reciprocal ranks across all models and metrics, providing a consolidated performance score that minimizes model-specific bias.

Ensemble learning with rotating downsampling and hyperparameter optimization

To maximize predictive performance while comprehensively utilizing available training data, we implemented an ensemble learning framework combining rotating downsampling with Bayesian hyperparameter optimization. Rather than performing a single random downsampling, we partitioned the majority class into 12 non-overlapping subsets, each matched 1:1 with the complete minority class (n = 127), generating 12 independent balanced training rounds through systematic rotation. This rotating strategy ensured that all available training data contributed to model learning while maintaining class balance in each round.

For hyperparameter optimization, six rounds were randomly selected from the 12 rounds to balance computational efficiency with robustness. For each algorithm (LR, SVM, RF, XGBoost, LightGBM), optimal hyperparameters were identified using Bayesian optimization (Optuna, 30 trials). Each trial’s performance was evaluated using 5-fold stratified cross-validation on each of the six tuning rounds: the five validation F1-scores were averaged within each round, yielding six round-level scores, which were then averaged to produce a single trial score. This nested cross-validation across multiple balanced rounds ensured that hyperparameter selection was robust to sampling variability. Hyperparameters maximizing this aggregated F1-score were retained as optimal. The complete hyperparameter search spaces are provided in Supplementary Table S1.

Using the optimized hyperparameters, final ensemble models were constructed by training individual sub-models on all 12 downsampling rounds. Ensemble predictions were generated by averaging the predicted probabilities across all sub-models, and final class labels were assigned using a probability threshold of 0.5. Performance was evaluated on the 2023 internal test set and 2024 temporal validation set. To evaluate the robustness of the class-imbalance handling strategy, additional sensitivity analyses comparing rotating downsampling, class-weighted learning, and SMOTE oversampling were conducted using the same feature set, data split, and tuned hyperparameters of the best-performing model.

Model interpretability and feature attribution

Model interpretability was assessed using SHAP, a game-theoretic approach to quantify feature contributions. For our ensemble framework, SHAP values were calculated individually for each sub-model trained on the rotating downsampled rounds and subsequently averaged to derive a robust global estimate of feature importance. We utilized beeswarm plots to visualize the distribution of SHAP values. To assess the stability of feature importance across the 12 rotating downsampling rounds, mean absolute SHAP values, standard deviations, coefficients of variation, and 95% confidence intervals were calculated for each feature. Detailed SHAP stability metrics are provided in Supplementary Table S2.

Epitope prediction and peptide-HLA interaction analysis of model-associated mutations

Candidate mutations selected from the top-ranked SHAP features associated with disease progression were subjected to epitope prediction and HLA binding affinity evaluation. HLA class I allele frequency data were obtained from the Allele Frequency Net Database (AFND, http://www.allelefrequencies.net/) using the “Chinese Jiangsu Han population” dataset, and alleles with a frequency > 0.1 were retained [16]. The distribution of mutation types within the cohort was summarized, and the most prevalent variants were selected for downstream analyses. For each mutation, wild-type and mutant protein sequences were constructed, and 9-mer and 10-mer candidate peptides spanning the mutation site were generated. Binding affinities between peptides and HLA class I molecules were predicted using NetMHCpan 4.2 (https://services.healthtech.dtu.dk/services/NetMHCpan-4.2/), with a selection threshold of %Rank ≤ 2.0 [17]. Peptides meeting the binding threshold were further evaluated for antigenicity using VaxiJen v2.0 (http://www.ddg-pharmfac.net/vaxijen/) with a cutoff ≥ 0.4 [18]. Predicted immunogenicity was assessed using the IEDB Class I Immunogenicity tool (http://tools.iedb.org/immunogenicity/), where positive scores indicate a higher potential for T-cell response [19]. Docking between HLA molecules and candidate peptides was performed using GalaxyPepDock [20] (http://galaxy.seoklab.org/cgi-bin/submit.cgi?type=PEPDOCK), and the binding free energy (ΔG) and dissociation constant (Kd) were calculated based on interface contacts (ICs) using PRODIGY (https://wenmr.science.uu.nl/prodigy/) [21]. The resulting HLA-peptide complexes were further visualized using PyMOL (version 3.0.3) to examine three-dimensional structural conformations and polar intermolecular interactions.

Results

Study cohort characteristics

Following quality control and feature selection, the final cohort comprised 5,072 samples with 45 features (Supplementary Table S3): 2,240 from 2023 (2,071 non-severe and 169 severe cases; 7.5% severe rate) and 2,832 from 2024 (2,754 non-severe and 78 severe cases; 2.8% severe rate). The 2023 cohort was partitioned into training (75%, n = 1,680; 127 severe cases) and internal test (25%, n = 560; 42 severe cases) sets for model development and initial validation. The 2024 cohort served as an independent temporal validation set to assess model generalizability across evolving viral variants. Additional year-stratified baseline characteristics, including age group, sex, vaccination status, prior infection history, regional distribution, and clinical severity categories, are provided in Supplementary Table S4. Viral sublineage composition shifted substantially between years: the 2023 cohort was dominated by XBB/EG.5- and HK.3/GF-related lineages, which were progressively replaced by JN.1- and XDV-related lineages during 2024 (Supplementary Fig. S1). Accordingly, the S: L452W and M: D3H substitutions were rare in 2023 but became near-universal across 2024 sublineages (Supplementary Table S5). Notably, this increase was not accompanied by a corresponding rise in the severe-case rate (Supplementary Fig. S2, Supplementary Table S6; see Supplementary Results), indicating that their temporal expansion primarily reflected lineage replacement rather than a direct increase in disease severity.

Systematic evaluation of feature combinations

To identify the optimal feature combination for clinical progression prediction, we systematically evaluated seven feature sets using five machine learning algorithms on the 2023 internal test set and 2024 temporal validation set. Performance was summarized using MRR calculated across all algorithms and metrics (F1-score, AUC, AUPR, accuracy) to provide an algorithm-agnostic ranking. On the 2023 test set, the comprehensive “All” feature set achieved the highest MRR (1.00), consistently outperforming all other combinations (Fig. 2; Table 1). Integrated feature sets substantially outperformed single-domain features, with “Clinical+Mutation” (MRR = 0.38), “Clinical+Region” (MRR = 0.33), and “Region” (MRR = 0.28) ranking next, indicating complementary predictive value across clinical, geographic, and genomic dimensions. Among individual algorithms using the “All” feature set, Random Forest achieved the best performance (F1 = 0.791, AUC = 0.819), followed by XGBoost (F1 = 0.780, AUC = 0.792) and LightGBM (F1 = 0.756, AUC = 0.799).

Fig. 2.

Performance comparison of seven feature combinations across machine learning models. A Results for the 2023 internal test set. B Results for the independent 2024 validation set. Rows indicate feature combinations, and columns indicate evaluation metrics. Bubble size represents the normalized performance score, bubble color represents metric-specific ranking, and numbers inside bubbles indicate the top-three ranks. Horizontal bars indicate the overall mean reciprocal rank (MRR) across the evaluated metrics

Table 1.

Performance of “All” feature set on the 2023 internal test cohort for COVID-19 severity prediction

Model	F1	AUC	AUPR	ACC
LR	0.714	0.800	0.758	0.714
SVM	0.709	0.776	0.752	0.726
XGBoost	0.780	0.792	0.781	0.786
LightGBM	0.756	0.799	0.778	0.762
RF	0.791	0.819	0.790	0.786

On the 2024 validation set, the “All” feature set maintained superior performance (MRR = 0.54), although absolute metrics declined, reflecting increased prediction difficulty due to class imbalance and temporal distribution shifts associated with ongoing viral evolution (Fig. 2; Table 2). Feature domain importance shifted slightly: “Clinical+Mutation” ranked second (MRR = 0.49), followed by “Clinical” (MRR = 0.46) and “Region” (MRR = 0.43). Despite these variations, the comprehensive “All” feature set consistently outperformed other combinations, suggesting that combining signals from multiple data domains supports temporal generalizability. Based on this systematic evaluation, the “All” feature set (n = 45; Supplementary Table S3) was selected for subsequent ensemble optimization and model interpretability analyses.

Table 2.

Performance of “All” feature set on the 2024 validation cohort for COVID-19 severity prediction

Model	F1	AUC	AUPR	ACC
LR	0.529	0.746	0.563	0.444
SVM	0.515	0.658	0.521	0.389
XGBoost	0.430	0.664	0.467	0.637
LightGBM	0.497	0.667	0.463	0.620
RF	0.494	0.653	0.433	0.632

Model performance evaluation

Following identification of the “All” feature set as optimal, five machine learning algorithms were trained with Bayesian optimized hyperparameters and evaluated on both temporal cohorts (Fig. 3; Table 3). On the 2023 internal test set, models demonstrated strong and comparable performance across discrimination metrics. For F1-score, the primary optimization objective balancing sensitivity and precision, SVM and LR tied for first place (both 0.713), closely followed by LightGBM (0.707). SVM achieved the highest discrimination (AUC = 0.786, AUPR = 0.805), slightly outperforming XGBoost (AUC = 0.781, AUPR = 0.796) and other models. LightGBM showed the highest positive predictive values (PPV = 0.725) and accuracy (0.714), indicating improved specificity in identifying true progression cases, a clinically relevant consideration for resource allocation in healthcare settings.

Fig. 3.

Comparative performance of machine learning models using the optimal “All” feature set. A Receiver operating characteristic (ROC) curves. B Radar plot summarizing key performance metrics across models

Table 3.

Performance of five machine learning algorithms trained with Bayesian-optimized hyperparameters using the “All” feature set

Metric	LR	SVM	RF	XGBoost	LightGBM
Test23_F1	0.713	0.713	0.643	0.667	0.707
Test23_AUC	0.753	0.786	0.753	0.781	0.752
Test23_AUPR	0.740	0.805	0.796	0.797	0.789
Test23_ACC	0.702	0.702	0.643	0.679	0.714
Test23_NPV	0.718	0.718	0.643	0.667	0.705
Test23_PPV	0.689	0.689	0.643	0.692	0.725
Test23_Brier	0.203	0.195	0.214	0.198	0.217
Test23_ECE	0.062	0.067	0.126	0.104	0.153
Test23_CalSlope	0.780	1.294	1.463	1.320	1.639
Test24_F1	0.566	0.586	0.568	0.593	0.603
Test24_AUC	0.720	0.750	0.742	0.748	0.735
Test24_AUPR	0.519	0.594	0.547	0.527	0.542
Test24_ACC	0.607	0.590	0.513	0.654	0.607
Test24_NPV	0.820	0.875	0.938	0.832	0.900
Test24_PPV	0.448	0.442	0.403	0.488	0.455
Test24_Brier	0.239	0.239	0.258	0.223	0.239
Test24_ECE	0.210	0.221	0.235	0.166	0.195
Test24_CalSlope	1.237	1.830	2.752	2.457	2.564

When evaluated on the 2024 temporal validation cohort, most models exhibited a decline in threshold-dependent metrics. Among the evaluated models, LightGBM achieved the highest F1-score (0.603), followed by XGBoost (0.593) and SVM (0.586), indicating relatively balanced sensitivity and precision under evolving epidemiological conditions. SVM showed marginally higher discrimination (AUC = 0.750; AUPR = 0.594), while other models exhibited comparable performance with only minor differences. Accuracy remained relatively similar across models, with XGBoost achieving the highest value (0.654), while LightGBM and LR showed comparable performance (0.607).

Calibration metrics were also evaluated for all models (Table 3). In the independent 2024 validation cohort, the LightGBM model showed a Brier score of 0.239, an ECE of 0.195, and a calibration slope of 2.564. Calibration slopes exceeded 1.0 across all five models in 2024, indicating imperfect probability calibration despite reasonable discrimination performance. These findings suggest that additional probability recalibration may be required before prospective clinical deployment.

Notably, all models exhibited high negative predictive values (NPV ≥ 0.82), with RF (0.938) and LightGBM (0.900) performing particularly well, indicating reliable identification of patients unlikely to progress. In contrast, PPV values were moderate across models (PPV = 0.40–0.49), likely reflecting the relatively low prevalence of severe outcomes in the dataset. Taken together, the combination of high NPV and moderate PPV suggests that the models may be particularly useful in clinical triage scenarios, where reliable exclusion of low-risk patients is prioritized to support efficient allocation of medical resources. Sensitivity analyses demonstrated that rotating downsampling achieved the strongest overall performance, particularly with respect to F1-score and AUC, in both the 2023 and 2024 cohorts compared with class-weighted learning and SMOTE oversampling (Supplementary Table S7).

SHAP-based feature importance and interaction analysis

The five algorithms were used for comparative benchmarking across different model families rather than as separate models intended for simultaneous deployment. Given that the primary objective was to identify predictors that generalize across time, LightGBM was selected for SHAP-based interpretation because it achieved the highest F1-score in the independent 2024 validation cohort (F1 = 0.603, Fig. 4). SHAP analysis revealed substantial heterogeneity in how host, viral, and clinical features contribute to COVID-19 progression risk (Fig. 4A). Age emerged as the dominant model predictor, with higher age values (indicated by purple) predominantly associated with positive SHAP distributions, indicating that advanced age substantially increases predicted severe outcome risk. The Age_Ct_N interaction term also ranked highly, suggesting potential synergistic effects between host susceptibility and virological parameters. Examination of the joint age-viral load distribution (Fig. 4B) revealed age-dependent modulation of viral load effects. Among individuals aged ≥ 60 years, SHAP values concentrated in the positive range with high magnitude. Notably, in elderly patients (particularly those aged 70–90 years), SHAP contributions remained elevated even at high Ct_N values (low viral load), suggesting that age-related model contribution remained substantial, whereas viral load exerts complex modulatory effects. In contrast, younger individuals at high Ct_N levels (low viral load) exhibited predominantly low or negative SHAP values, suggesting that age plays a critical regulatory role in how viral burden influences disease progression trajectories. Additional density-contour and stratified heatmap visualizations of the Age_Ct_N interaction are provided in Supplementary Fig. S3. To further support this interaction pattern, we stratified severe-case rates by age group and Ct_N category. Older age groups generally showed higher severe-case rates across Ct_N strata, including in the Ct_N ≥ 30 group, where severe-case rates were 10.7% and 8.7% among individuals aged 60 to 69 years and ≥ 70 years, respectively, compared with 2.3% among those aged < 60 years (Supplementary Table S8). To evaluate the robustness of feature attribution within the ensemble framework, SHAP importance stability was assessed across all 12 rotating downsampling rounds. Feature rankings were generally consistent, with Age and the Age_Ct_N interaction remaining the two most influential predictors across rounds. The corresponding mean round-level |SHAP| values, standard deviations, coefficients of variation, and approximate 95% confidence intervals are provided in Supplementary Table S2, and the stability of the top-ranked features is illustrated in Supplementary Fig. S4.

Fig. 4.

SHAP-based interpretation of the LightGBM model. A Beeswarm plot displays the top 20 features ranked by mean absolute SHAP value. Each point represents an individual prediction. The horizontal axis indicates SHAP value, reflecting the magnitude and direction of feature contribution to predicted severity risk. Color represents feature value (purple = high; green = low). B SHAP interaction scatter plot illustrating the joint effect of age and Ct_N values on predicted severity risk

Among geographic factors, the southern Jiangsu region (Region_SuNan) contributed substantially to model predictions, potentially reflecting regional differences in population structure, epidemiological context, or healthcare resource allocation. Vaccination status ranked as an important feature, demonstrating association with reduced severe outcome risk. Within viral genetic features, M protein mutation and ORF1a mutation count both exhibited high importance, with SHAP distributions showing bidirectional patterns that suggest nonlinear relationships between mutation burden and clinical outcomes. Among specific mutation sites, S.368 displayed predominantly negative SHAP contributions, potentially associated with reduced risk, while M.3 showed positive risk signals. Additionally, physicochemical mutation load metrics (Size_Change_Mut_Load, Charge_Change_Mut_Load) demonstrated moderate contributions, suggesting that amino acid property alterations may influence viral biological characteristics and disease severity.

Among clinical factors, prior infection history exhibited negative SHAP distributions, indicating protective effects, while gender showed male-associated elevated risk. Ct value metrics (Ct_ORF1ab, Ct_N) when analyzed independently showed relatively limited contributions with mixed SHAP distributions, suggesting that viral load effects on disease severity may depend on host background and synergistic interactions with other factors. Remaining mutation sites (e.g., E.11, S.183, S.490) and regional variables (Region_SuBei) exerted relatively minor influence with predominantly negative contributions. Although S.452 ranked lower overall, its high feature values corresponded to positive SHAP distributions, suggesting potential risk association.

Taken together, the SHAP results point to a layered risk structure: among the variables available to the model, host age showed the largest predictive contribution, viral genetic variation and vaccination status add biological modulation, and regional factors (particularly southern Jiangsu) capture differences in the broader epidemiological environment. Disease severity thus appears to be shaped by the interplay of host, virus, and contextual factors operating at different scales.

Population-level epitope presentation and structural characterization of severity-associated mutations

To investigate the potential immunological implications of severity-associated mutations identified by SHAP analysis (S.452 and M.3), epitope presentation potential was evaluated using several high-frequency HLA class I alleles in the Jiangsu population (HLA-A*02:01, A*11:01, A*24:02, A*30:01, A*33:03, and HLA-B*13:02) (Table 4). At the S protein position 452, the predominant mutation observed in the cohort was L452W (n = 2,795). Epitope prediction analysis showed that the wild-type L452 region generated two antigenic and immunogenic candidate peptides, including KVGGNYNYL (binding to HLA-A*02:01 and A*30:01) and KVGGNYNYLY (binding to HLA-A*11:01). In contrast, the L452W mutation retained only one antigenic and immunogenic peptide, KVGGNYNYW, which only bound to HLA-A*24:02. This reduction in the predicted epitope repertoire suggests a potential decrease in antigen presentation diversity.

Table 4.

Predicted epitope presentation and peptide–HLA binding parameters for severity-associated mutations

Peptide	HLA	Binding Level	Antigenicity	Immunogenicity	ΔG (kcal/mol)	Kd (M)
KVGGNYNYL	A*02:01	1.670WB	0.599	0.028	-10.400	2.5e-08
KVGGNYNYL	A*30:01	1.035WB	0.599	0.028	-10.100	4.0e-08
KVGGNYNYLY	A*11:01	1.715WB	0.534	0.020	/	/
KVGGNYNYW	A*24:02	1.040WB	0.534	0.028	-9.800	6.8e-08
ADSNGTITV	B*13:02	1.613WB	Non-antigen	0.144	/	/
AHSNGTITV	B*13:02	1.835WB	0.445	0.144	/	/

*WB Weak binder; binding level represents NetMHCpan %Rank

At the M protein position 3, the dominant mutation was D3H (n = 2,775). The wild-type peptide ADSNGTITV showed predicted binding to HLA-B*13:02 but lacked antigenicity. In contrast, the mutant peptide AHSNGTITV (D3H) was predicted to possess both antigenicity and immunogenicity while maintaining binding to HLA-B*13:02, although with slightly reduced binding affinity compared to the wild-type peptide. These findings suggest that M.3 may alter predicted epitope properties, but whether this translates into altered immune recognition or clinical risk requires experimental validation.

Due to the absence of an experimentally resolved structure for HLA-B*13:02 in the Protein Data Bank, structural docking analysis was limited to peptide-HLA complexes with available crystal structures. The wild-type peptide KVGGNYNYL was docked with HLA-A*02:01 and HLA-A*30:01, whereas the mutant peptide KVGGNYNYW was docked with HLA-A*24:02. Predicted binding stability metrics are summarized in Table 4. Although direct quantitative comparison should be interpreted cautiously because the peptides were evaluated in different HLA-restriction contexts, the docking results suggested potential mutation-associated changes in predicted peptide-HLA interaction patterns. Detailed template selection and structural interaction observations are provided in Supplementary Results and Supplementary Fig. S5.

Discussion

Drawing on real-world surveillance data from Jiangsu Province, this study combined host clinical characteristics, viral genetic variation, and regional epidemiological information to assess the risk of COVID-19 progression. Across feature combinations and algorithms, the integrated feature set consistently outperformed single-domain inputs, pointing to a multidimensional basis for COVID-19 severity. A rotating downsampling strategy paired with ensemble learning was used to address class imbalance while retaining as much training information as possible. The LightGBM model trained on the integrated feature set ultimately achieved an F1-score of 0.603 in the temporally independent 2024 validation cohort, indicating a degree of cross-temporal generalizability. SHAP analysis revealed a layered risk structure in which host susceptibility factors carried the greatest weight, viral genetic features added biological detail, and regional variables captured differences in the epidemiological and healthcare environment. During the Omicron circulation period, disease severity in our model was most strongly associated with host vulnerability-related factors, while viral genetic background and regional epidemiological context provided additional predictive information.

SHAP results identified age as the dominant predictor of disease severity, consistent with prior studies. Older individuals remained at elevated risk for severe outcomes during the Omicron period, potentially attributable to immunosenescence, increased burden of comorbidities, and diminished regulation of inflammatory responses [22, 23]. Notably, an interaction effect between age and Ct_N values was observed. The N gene was selected as a viral load indicator owing to its high expression in subgenomic RNA, superior detection sensitivity in low-viral-load samples, and relative sequence conservation, rendering it a suitable quantitative reference [22, 24]. Previous studies have reported inconsistent findings regarding the role of viral load in disease progression, likely influenced by sampling timing and host characteristics. Liu et al. observed that patients with severe COVID-19 tended to have a higher viral load and prolonged virus shedding [25]. In contrast, Cevik et al. reported that viral loads were similar at the start of infection between asymptomatic and symptomatic SARS-CoV-2 patients [26]. In this study, even among older individuals with relatively low viral loads, elevated severe-case rates and positive model contributions were still observed, suggesting that host vulnerability may play a dominant role in risk formation, whereas viral load may function more as a modulatory factor.

In addition to age, sex demonstrated predictive relevance, with male patients exhibiting increased risk of severe outcomes, consistent with multiple meta-analyses [27–29]. Vaccination history showed a protective contribution in SHAP analyses, supporting the role of vaccination in reducing severe disease risk associated with Omicron variants [30, 31]. Prior infection history also exhibited a degree of protective effect, potentially reflecting broader cross-reactive immunity conferred by hybrid immune responses [32].

The study further identified S.452 and M.3 sites as model features associated with disease severity. Epitope prediction analyses suggested that the S protein L452W mutation may reduce the number of potential predicted candidate epitopes within the evaluated Jiangsu high-frequency HLA allele set. Previous experimental studies have reported that mutations involving residue 452 may influence immune recognition and T-cell responses in certain SARS-CoV-2 variants, providing biological context for the present computational observations [33, 34]. Nevertheless, the biological significance of the L452W mutation remains to be experimentally validated. Structural docking analyses further suggested that mutation-associated peptides may exhibit altered peptide-HLA interaction patterns and slightly different predicted binding stability in the available HLA contexts. Because these comparisons involved different HLA restriction backgrounds, the predicted binding free energies should not be interpreted as direct quantitative evidence of reduced presentation efficiency. The M protein D3H mutation was predicted to alter epitope antigenicity, suggesting that non-spike structural proteins may also influence antigen presentation characteristics. However, these findings are based on computational predictions and require further functional validation.

Notably, regional factors also demonstrated measurable influence within the model, particularly variables associated with southern Jiangsu. These findings are more likely to reflect macro-epidemiological factors such as demographic structure, healthcare accessibility, vaccination strategies, and regional transmission dynamics rather than inherent geographic determinants of disease severity, highlighting the importance of incorporating population-level information into risk modeling.

Several limitations should be acknowledged. First, this retrospective analysis was conducted within a single province in eastern China, and external generalizability requires validation in multi-regional cohorts. In particular, the epitope prediction results may vary across populations with different HLA backgrounds. In addition, although the surveillance program covered all 13 prefecture-level cities in Jiangsu Province, the analyzed cohort may not fully represent all SARS-CoV-2 infections occurring during the study period because sequencing success depends on sample availability and quality. Furthermore, the SARS-CoV-2 evolutionary landscape continues to change rapidly, and some mutation-associated risk signals identified in this study may become less relevant as new variants emerge and population immunity profiles evolve. Continuous model updating and prospective validation will therefore be necessary to maintain predictive performance under future endemic conditions. Second, detailed information on chronic comorbidities (e.g., hypertension, cardiovascular disease, chronic respiratory disease, diabetes, and immunodeficiency conditions) was not available in the surveillance dataset. In addition, individual immunological indicators, including neutralizing antibody titers and cellular immune status, were not incorporated. These limitations prevented direct quantification of the relative contributions of underlying disease burden and hybrid immunity to disease severity. Additionally, vaccination data were not stratified by specific vaccine types or dosing intervals. Interactions between vaccination status and specific viral mutations were also not explicitly modeled because introducing numerous sparse interaction terms would have increased model dimensionality relative to the number of severe cases. From a methodological perspective, although the rotating downsampling combined with ensemble learning helped mitigate information loss, this strategy may still influence probability calibration and model generalizability in real-world populations. Finally, epitope and structural predictions were based on computational modeling and require further immunological and functional validation.

Overall, this study suggests that COVID-19 disease severity during the endemic phase is shaped by multiple hierarchical layers, including host susceptibility, viral genetic variation, and macro-level epidemiological context. The proposed multimodal machine learning framework may provide a useful analytical paradigm for integrating genomic surveillance with clinical prediction in COVID-19 and other emerging infectious diseases. With appropriate external validation, model updating, and probability recalibration, such approaches may ultimately support the development of user-friendly decision-support tools for real-time risk assessment and public health surveillance.

Conclusion

Using real-world surveillance data from Jiangsu Province, we developed and validated severity prediction models integrating clinical, genomic, and regional features, with LightGBM achieving the strongest temporal validation performance (F1-score: 0.603). SHAP-based interpretation identified age as the strongest model predictor, while vaccination status, prior infection history, and selected viral genomic features provided additional predictive information. Bioinformatics analyses of the model-associated S.452 and M.3 sites suggested potential changes in predicted peptide-HLA interactions; as these findings are computational predictions, experimental validation is required before their biological significance can be established. Overall, these findings support the feasibility of integrating multidimensional surveillance data for severity risk prediction and, with appropriate external validation and probability recalibration, may inform clinical triage and public health surveillance of emerging variants.

Authors’ contributions

X.Y.M. drafted the original manuscript. X.Y.M. and S.D.F. performed the data analysis. W.R.X., Y.S., and L.Q.W. conducted the investigation and validation. X.L. and H.Y.Y. reviewed and edited the manuscript. J.L. and L.G.Z. conceived and designed the study and supervised the conceptualization process. All authors reviewed the manuscript.

Funding

This work was supported by the CSS Program (HYZHXMH01001); National Key R&D Program of China (2025YFF0512800); the National Natural Science Foundation of China (32270607); the Social Development Foundation of Jiangsu Province (BE2021739, BE2023601), and the Scientific research project of Jiangsu health commission (DX202301); State Key Laboratory of Space Medicine, China Astronaut Research and Training Center (SKL 2024K03); the Open Research Fund of Southeast University and Jiangsu Province Hospital; Jiangsu Province 333 project. The funding bodies were not involved in the study design, data collection, data analysis and data interpretation.

Data availability

The SARS-CoV-2 whole-genome sequences analyzed in this study were generated under the routine public health surveillance conducted by the Jiangsu Provincial Center for Disease Control and Prevention (Jiangsu CDC) and were submitted to the National Pathogen Vector and Host Animal Surveillance Management Information System (NPVMS, https://npvms.icdc.cn/), a centralized data management platform maintained by the Chinese Center for Disease Control and Prevention (China CDC) for routine infectious disease genomic surveillance. Due to the sensitive nature of public health surveillance data, the sequences and the associated clinical metadata are available from us upon reasonable request. The analysis code is publicly available at https://github.com/mmsnowy/COVID-Severity-Prediction.

Declarations

Ethics approval and consent to participate

The study received ethical exemption approval from the Ethics Committee of Jiangsu Provincial Center for Disease Control and Prevention. The requirement for informed consent was waived by the Ethics Committee of Jiangsu Provincial Center for Disease Control and Prevention. The SARS-CoV-2 surveillance data used in this study were collected as part of statutorily required infectious disease prevention and control measures. These data were obtained from legitimate sources and stored securely, and their secondary use fully adhered to ethical standards. All statistical analyses were conducted using de-identified datasets, thereby guaranteeing that the rights and interests of study participants were fully protected and no potential harm or rights violation was incurred. Given that the study is beneficial to COVID-19 prevention and control and presents minimal risk, it meets the criteria for exemption from ethical review. All methods were performed in accordance with the relevant guidelines and regulations, including the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.