According to the latest report from the China Center for Disease Control and Prevention, as of December 31, 2024, there were 1,355,017 people living with human immunodeficiency virus (HIV)/acquired immunodeficiency syndrome (AIDS) in China, with 491,437 cumulative reported deaths (1). Current first-line antiretroviral therapy (ART) in China, typically consisting of nucleoside reverse transcriptase inhibitors (NRTIs) combined with non-nucleoside reverse transcriptase inhibitors (NNRTIs), protease inhibitors (PIs), or integrase strand transfer inhibitors (INSTIs), effectively suppresses viral replication (2). With the widespread application of ART, the viral load suppression rate among people living with HIV (PLWH) in China had risen to 95.4% by the end of 2021 (3). However, despite the significant overall effectiveness of AIDS antiviral therapy and the continuous decline in the population mortality rate, there remains a subset of patients who, despite achieving virological suppression (viral load < 50 copies/mL), experience difficulty in restoring their CD4⁺ T lymphocyte counts to a relatively satisfactory immune function level (≥350 cells/μL). This population is clinically defined as immunological non-responders (INRs) (4). INRs face a substantially higher risk of mortality and opportunistic infections compared to immunological responders (IRs), with non-AIDS-defining events becoming their leading cause of death, severely impacting long-term quality of life (5, 6).

In recent years, immune non-response (INR) has attracted increasing clinical attention, and its occurrence is closely associated with multiple risk factors. Previous studies have shown that advanced age and low pre-treatment CD4⁺ T cell count are major risk factors for INR (7–10). Other reported factors include the route of infection (11), ART regimen (12–14), delayed ART initiation (15), and World Health Organization (WHO) clinical stage (16) are also associated with poor prognosis in PLWH; however, the specific roles of these variables in immune reconstitution assessment remain controversial. Various laboratory parameters, such as white blood cell count (WBC) (17), platelet count (PLT) (18, 19), hemoglobin (Hb) (20, 21), alanine aminotransferase (ALT) (22), aspartate aminotransferase (AST) (16), and serum creatinine (SCr) (23), have also been linked to HIV disease progression and mortality. Furthermore, viral coinfections (e.g., hepatitis B and C) have been confirmed to impede immune recovery (24–26). The pathogenesis of INR has not been fully elucidated. Existing studies suggest that it is associated with an imbalance between T lymphocyte production and apoptosis caused by insufficient thymic output (27, 28), chronic immune activation and T-cell exhaustion (29, 30), intestinal flora disturbance and microbial translocation (31), as well as metabolic abnormalities such as ferroptosis and mitochondrial dysfunction in CD4⁺ T cells (32).

Following a systematic review of the existing literature on prediction models for INR in PLWH, we identified several limitations in current research. First, there is a lack of uniformity in the definition and diagnostic criteria for INR across different studies. This leads to heterogeneity in the outcome variables targeted by the prediction models, making direct comparison and integration of findings difficult and consequently limiting their clinical relevance (17). Second, regarding model development, existing studies have predominantly focused on predicting the risk of early INR after treatment initiation. For instance, although the study by Zhang et al. (16) incorporated a broad set of variables to enhance predictive comprehensiveness, some indicators (e.g., specific infection history) are not routinely available in clinical practice, which may compromise the model’s utility. Furthermore, the prediction time window of this model is primarily confined to the initial treatment phase (e.g., 2 years after treatment). Key performance metrics, such as discrimination and calibration, have not been adequately validated in patients receiving long-term virologically suppressive therapy. Therefore, its applicability for guiding INR management in individuals on extended treatment (e.g., beyond 4 years) remains limited.

Although China issued the consensus on diagnosis and management of immunological non-responders in HIV infection (Version 2023) in recent years to standardize diagnostic criteria (4), prediction models based on this consensus remain scarce. To fill this research gap, this study strictly adheres to the above consensus standards to develop and validate a risk prediction model for INR.

Machine learning algorithms have shown significant promise in disease diagnosis and prognosis prediction (33). As an intuitive risk visualization tool, the nomogram can integrate multiple predictive variables into an evaluation model for the probability of clinical events based on the results of multivariable regression analysis. Therefore, to improve the accuracy and clinical utility of INR risk identification, this study combines machine learning-based variable selection with nomogram model construction, aiming to provide an effective tool for the early identification of high-risk INR individuals for targeted intervention.

The significantly elevated risk of adverse events in INRs creates an urgent clinical need for early identification of high-risk individuals. To address the long-standing core issue of inconsistent diagnostic criteria in this research area, this study strictly defined outcome events according to the latest Chinese expert consensus, aiming to provide more reliable evidence for this field.

This study identified age at ART initiation, pre-treatment CD4⁺ T cell count, and AST level as independent predictors of INR and constructed a predictive model based on these three factors, which is consistent with numerous previous reports (17, 21, 34). Existing studies have shown a positive correlation between age at ART initiation and INR risk (8, 35), while a lower pre-treatment CD4⁺ T cell count is a strong predictor, demonstrating a significant negative correlation with INR occurrence (36, 37). Further research by Li et al. revealed that the CD4⁺ T cell count in individuals over 50 years old was significantly lower than in those under 35. Although ART helped restore CD4⁺ T cell counts to some extent across all age groups, younger patients exhibited significantly superior immune reconstitution rates and degrees of recovery compared to older patients. Moreover, the inhibitory effect of age on CD4⁺ T cell recovery became increasingly pronounced with longer treatment duration (38). Mechanistically, the thymus, as the primary site for generating immunocompetent T cells, undergoes progressive functional decline with age, leading to restricted nascent T cell output and consequently impairing CD4⁺ T cell recovery (39). Additionally, elderly patients have shortened T cell telomere lengths, weakened proliferation ability, and are prone to immunosenescence (40). This mechanistic understanding supports the negative impact of advanced age on INR observed in our study. Research indicates that advanced age and low CD4⁺ T cell count are major risk factors for significant clinical events such as cardiovascular events and non-AIDS-defining malignancies (5). Another study reported that the median age of the high-risk group for atherosclerotic cardiovascular disease among PLWH was 66 years, significantly higher than the 52 years in the low-risk group (41), underscoring the critical role of age in non-AIDS-related complications. Furthermore, a study observing immune recovery after one year of ART noted that patients with a baseline CD4⁺ T cell count <50/μL struggled to achieve a CD4⁺ T cell count above 200/μL despite standardized ART (42). Another study also emphasized that the baseline CD4⁺ T cell count is a sensitive indicator for predicting immune reconstitution outcomes (43).

AST serves as a crucial biomarker for liver function assessment. Relevant studies indicate that AST levels are closely and positively correlated with viral suppression and immune recovery in PLWH (16). As a sensitive indicator of hepatocellular damage, elevated AST levels reflect hepatic inflammation or necrosis. Impaired liver function subsequently reduces the capacity to clear toxins and inflammatory factors, leading to a systemic chronic low-grade inflammatory state and persistent immune activation. This pathological environment can accelerate functional exhaustion of CD4⁺ T cells, thereby impeding the immune reconstitution process (44). In our study, patients with AST levels >40 U/L exhibited a 2.60-fold higher risk of developing INR compared to those with AST ≤40 U/L, further supporting the value of incorporating AST levels as a predictive indicator for INR occurrence.

This study addresses the clinically relevant issue of CD4⁺ T cell recovery plateau following long-term viral suppression within the context of China’s predominant ART regimens, which primarily consist of two NRTIs combined with either one NNRTI, one PI, or one INSTI (45). This focus underscores the study’s significant practical relevance and research necessity. By tracking annual CD4⁺ T cell count changes from treatment initiation until the follow-up cutoff, and applying the INR diagnostic criteria recommended by the world’s first expert consensus on INR, we conducted a retrospective analysis of pre-ART clinical data from 615 PLWH at a single tertiary hospital in a major central Chinese city. We identified age, CD4⁺ T cell count, and AST level as baseline risk factors for INR and pioneered the development of a predictive nomogram model based on these three parameters. External validation was performed using a follow-up cohort from a primary hospital.

The model demonstrated exceptional discrimination in both the training set (AUC = 0.896) and internal validation set (AUC = 0.903), indicating excellent predictive performance. Notably, in the independent external validation set, the model achieved an AUC of 0.766, exceeding the empirical threshold of 0.75, which suggests reasonable generalizability. Furthermore, calibration curves and decision curve analysis confirmed the model’s satisfactory calibration and clinical applicability. This study facilitates early risk stratification and tailored follow-up management for PLWH at high risk of INR, ultimately aiming to improve patient prognosis and quality of life.

Furthermore, compared to the relatively lenient criteria used in earlier studies (46), the updated standards provide a more precise definition of INR. Previous criteria often suffered from limitations such as short observation periods and inconsistent thresholds. In contrast, the new standards not only require a sustained CD4⁺ T-cell count below a specific threshold but also emphasize evaluation under the condition of prolonged viral suppression (e.g., >3 years). This approach more accurately identifies the true “inadequate immune reconstitution” population and reduces misclassification due to incomplete viral suppression or other confounding factors. Consequently, the predictors identified in this context—such as baseline CD4⁺ T-cell count, AST level, and age—carry greater biological relevance and robustness.

In terms of clinical interpretation, models based on the updated criteria place greater emphasis on identifying high-risk individuals. Older criteria may have included many patients with relatively low long-term risk, whereas the predicted probabilities from our model better reflect a patient’s actual risk of long-term adverse outcomes, thereby offering a more reliable basis for clinical decision-making.

More importantly, the evolution of diagnostic consensus reflects a deepened understanding of disease prognosis management. Previously, inconsistent diagnostic criteria often led to unclear indications and necessity for clinical intervention. The predictive model constructed according to the new standards enables risk stratification and facilitates the early identification of patients who, despite successful virological control, remain in an immunodeficient state and thus require closer monitoring or consideration of adjunctive immune interventions. For example, in patients classified as high-risk by the model, clinicians may consider shortening follow-up intervals, enhancing screening for opportunistic infections, or more actively exploring opportunities for participation in immunotherapy clinical trials.

Several limitations warrant consideration. First, a degradation in model performance was observed in the external validation cohort, as indicated by a decrease in the AUC values. The observed decline in performance in the external validation set may stem from significant demographic and clinical differences between the tertiary and primary hospital cohorts. Specifically, notable disparities were identified: the male-to-female ratio was approximately 3:1 in the external cohort compared to about 19:1 in the training cohort. Furthermore, heterosexual transmission was predominant in the external cohort, whereas homosexual transmission was more common in the training cohort. Patients in the external cohort also had a higher overall age at ART initiation. In addition, marked differences existed in baseline laboratory indicators (e.g., WBC, PLT, Hb) between the two cohorts.

These discrepancies may affect the model’s generalizability through several mechanisms. Firstly, baseline immune status and laboratory profiles likely exhibit population-specific variations across different genders, transmission routes, and age groups. Consequently, a model developed on the training cohort may naturally show reduced predictive accuracy when applied to an external population with a different feature distribution. Secondly, these phenotypic characteristics could act as proxies for deeper, unmeasured biological factors—such as viral subtypes, specific inflammatory biomarkers, or host genetic variations—that may influence INR, potentially explaining part of the performance gap (29, 47). Thirdly, since our analysis confirmed that age at ART initiation is an independent predictor of incomplete immune reconstitution, the difference in age distribution between cohorts could directly impact the model’s calibration and discriminative ability. Finally, the retrospective design is susceptible to selection and attrition bias.

Future studies should aim to enhance the model’s robustness by expanding the sample size through the inclusion of patient cohorts from more diverse geographic regions and ethnic backgrounds. The consideration of more advanced machine learning algorithms may also be warranted to better adapt the tool to various clinical scenarios.

In summary, we have constructed a practical nomogram that effectively predicts the risk of incomplete immune reconstitution in virologically suppressed PLWH. By integrating age at ART initiation, baseline CD4⁺ T cell count, and AST level, this tool allows for early risk stratification. It can assist clinicians in identifying high-risk INR individuals who may benefit from more intensive monitoring, investigation of comorbidities, and personalized management strategies, ultimately aiming to improve long-term clinical outcomes.