Osteosarcoma (OS) is a primary malignant bone tumor characterized by the production of osteoid tissue, most commonly occurring in the metaphysis of long bones in adolescents aged 10–20 years (1). Prior to the 1970s, treatment relied primarily on surgery alone, with poor outcomes; up to 80% of patients ultimately died from pulmonary metastases, and the 5-year overall survival (OS) rate was less than 20%. The landscape changed significantly by the 1980s. A comprehensive treatment protocol incorporating preoperative chemotherapy, surgery, and postoperative chemotherapy was established as the standard of care. This multimodal approach increased the 5-year OS rate to over 60% for patients without metastases at initial diagnosis (2, 3). However, the management of OS continues to pose significant challenges. Approximately 15% of patients present with metastases at diagnosis, conferring a very poor prognosis. Furthermore, about 30% of patients without metastases initially may develop metastases during treatment, leading to a substantial decline in survival (4). While novel therapies, such as targeted treatments and immunotherapy, have been applied for OS in recent years, their efficacy in improving outcomes for patients with metastatic disease remains limited (5, 6). Consequently, the Children’s Oncology Group (COG) suggests that the early identification of high-risk patients and the development of individualized treatment plans may represent another crucial pathway for improving prognosis (7).

Lymphocytes, as core components of the immune system, play an increasingly recognized role in tumor initiation, progression, and control. Multiple studies have shown that the presence of tumor-infiltrating lymphocytes (TILs), particularly CD8+ T cells, in tumor tissue is generally associated with a more favorable prognosis (8, 9). Similarly, biomarkers derived from peripheral blood lymphocyte counts, such as the neutrophil-to-lymphocyte ratio (NLR), have been established as independent prognostic factors for various solid tumors, including OS, with a lower pre-treatment NLR often correlating with better survival outcomes (10–12). Advances in lymphocyte research have revealed the complex and often opposing functions of different lymphocyte subsets in tumor progression. For instance, while CD8+ T cells can directly kill tumor cells, regulatory T cells (Tregs) may suppress immune responses and potentially promote tumor growth (13, 14). However, conventional blood tests typically report only the total lymphocyte count and cannot differentiate between subsets such as T cells, B cells, and natural killer (NK) cells. Therefore, peripheral blood lymphocyte subset enumeration holds significant potential. It can provide a more comprehensive reflection of a patient’s systemic immune status, enabling a more authentic and objective assessment of overall immune activity. This is crucial for precise prognosis identification and guiding treatment strategies (15).

Although peripheral blood lymphocyte subset analysis can simultaneously measure over ten lymphocyte subsets in a patient’s blood, providing rich data for assessing systemic immune status, the processing and interpretation of the resulting multidimensional and complex data remain a significant challenge (16). In this context, machine learning (ML), as a powerful artificial intelligence (AI) method, has been demonstrated to significantly enhance the analysis and interpretation of complex cancer datasets (17, 18). A key advantage of ML methods is their ability to utilize relatively limited data to perform in-depth analysis of the complex biology of diseases and achieve personalized risk stratification, thereby holding promise for improving patient treatment outcomes (19, 20). Despite this clear potential, the application of ML for the deep analysis of OS-specific lymphocyte subset data has not yet been fully explored. This gap extends to the subsequent development of novel biomarkers. Such an endeavor highlights substantial research potential and value in this field.

Based on the aforementioned background, this study aims to construct a prognostic prediction model by integrating peripheral blood lymphocyte subset data with clinicopathological characteristics of OS patients using an ML algorithm. This research seeks to systematically evaluate the independent and synergistic prognostic value of different lymphocyte subsets and explore their combined utility with established clinical indicators. Ultimately, we intend to develop an effective tool capable of accurately identifying high-risk patients and informing individualized treatment strategies, thereby contributing to the advancement of precision medicine practice.

Through a systematic analysis of peripheral blood lymphocyte subsets in 65 OS patients, this study revealed significant associations between immune characteristics of lymphocytes and patient prognosis. It was found that the distribution of lymphocyte subsets differed significantly between deceased and surviving patient groups. Specifically, the proportions of six subsets—CD3⁺CD8⁺, CD3^-CD56⁺, CD4⁺CD28⁺, CD8⁺CD28⁺, CD4⁺HLA-DR⁺, and CD8⁺HLA-DR⁺—were significantly lower in the deceased group compared to the survival group. Notably, although an increasing trend was observed in the proportions of the CD4⁺CD38⁺ and CD8⁺CD38⁺ subsets within the deceased group, this difference did not reach statistical significance. By comparing various ML algorithms, a prognostic prediction model based on the GBM was established. This model, utilizing only two variables—CD3^-CD56⁺ NK cells and CD8⁺HLA-DR⁺ activated cytotoxic T cells—achieved excellent predictive performance, with an AUC of 0.9591. Furthermore, the gbm_risk_score, constructed based on these two lymphocyte subsets, showed a significant negative correlation with overall patient survival and was confirmed as an independent prognostic factor in multivariate analysis, distinct from traditional clinical factors such as metastasis status at initial diagnosis. The subsequently developed nomogram model demonstrated good accuracy and clinical utility in predicting the 3-year OS rate, successfully identifying high-risk subgroups among patients without initial metastasis. These findings provide new perspectives and methodologies for the immune prognostic assessment of OS.

Previous studies have extensively explored the role of total peripheral blood lymphocyte count and its derived biomarkers, such as the NLR, in prognostic assessment across various cancers (21–23). For instance, in extensive-stage small cell lung cancer (ES-SCLC), research has demonstrated that patients with a low pretreatment total lymphocyte count or a high NLR exhibit shorter OS (24). Similarly, a meta-analysis focusing on gastric cancer (GC) patients revealed that elevated NLR and platelet-to-lymphocyte ratio (PLR) were associated with poorer OS and progression-free survival (PFS) following treatment with immune checkpoint inhibitors (ICIs), whereas a high lymphocyte-to-monocyte ratio (LMR) was correlated with improved survival outcomes (25). Furthermore, in OS research, NLR has been confirmed as an independent risk factor for predicting patient survival, and dynamic monitoring of NLR changes can further enhance the accuracy of prognostic evaluation (10). Collectively, these findings underscore the significant value of total peripheral blood lymphocyte count and its derived ratios in pan-cancer prognostic prediction and even in the evaluation of treatment response.

With the deepening of research on the tumor immune microenvironment (TIME), the functional diversity of lymphocyte subsets and their complex roles in tumorigenesis and development have become increasingly clear (26). Among them, CD4+ T cells are crucial for initiating and regulating adaptive immunity; they effectively promote anti-tumor immune responses by supporting the clonal expansion, differentiation, and memory formation of cytotoxic T lymphocytes (CTLs) (27, 28). CD8+ T cells, as direct effector cells, kill tumor cells through the release of cytotoxic substances such as perforin and granzymes (29). In contrast, Tregs play an immunosuppressive role in the TIME by inhibiting the function of effector T cells, thereby promoting tumor immune escape and progression (30); studies have shown that the proportion of Tregs in the peripheral blood of some cancer patients is significantly higher than that in healthy individuals (31). Therefore, compared to measuring the total lymphocyte count alone, precise analysis of these functionally diverse and often antagonistic lymphocyte subsets in peripheral blood provides a more comprehensive reflection of the overall immune status of cancer patients and demonstrates greater potential for clinical application in prognostic prediction.

CD56+ NK cells, one of the cornerstones of the gbm_risk_score, constitute a central component of innate anti-tumor immunity, which can directly lyse tumor cells without the need for prior activation, notably those with low or absent MHC-I expression (32). Accumulating evidence indicates that the presence of NK cells is generally associated with a favorable prognosis. For instance, in patients with advanced hepatocellular carcinoma treated with ICIs, an increased proportion of peripheral blood NK cells at week 3 post-treatment is an independent predictor of objective response and long-term survival (PFS and OS) (33). Similarly, in advanced gastric cancer patients treated with apatinib, a peripheral blood NK cell proportion below 17% is associated with poorer PFS and OS (34). In osteosarcoma, a dual-center retrospective study of 106 patients demonstrated that a high proportion of NK cells is correlated with improved survival (35). Furthermore, preclinical strategies aimed at enhancing NK cell migration and infiltration into osteosarcoma tumors have yielded promising preliminary results (36). These findings align with our own data, thereby highlighting the critical role of NK cells in tumor immune surveillance via direct cytotoxicity.

CD8⁺HLA-DR⁺ T cells represent a subset of CTLs commonly regarded as markers of activated, effector memory T cells (37). HLA-DR is a class II MHC molecule whose expression on T cells typically denotes an activated state associated with antigen presentation or effector functions. Existing research suggests that CD8⁺HLA-DR⁺ T cells can have dual prognostic significance in cancer patients. For example, studies in breast cancer show that high levels of HLA-DR⁺ CTLs are associated with a favorable response to neoadjuvant chemotherapy and improved PFS, indicating a positive prognostic role (38, 39). However, other research indicates that in primary central nervous system lymphoma, elevated levels of CD8⁺HLA-DR⁺ T cells has been linked to T cell exhaustion and an immunosuppressive microenvironment, suggesting a poor prognosis (37). In contrast, the prognostic value of CD8⁺HLA-DR⁺ T cells in osteosarcoma has not been definitively established. Only one study has proposed that increased HLA-DR expression in osteosarcoma likely reflects an ongoing anti-tumor immune response (40). In our study, a higher proportion of peripheral blood CD8⁺HLA-DR⁺ T cells was associated with prolonged OS. This implies that within the context of osteosarcoma, this cell population may represent an effective anti-tumor immune response rather than an exhausted phenotype. Further studies are warranted to validate this finding.

While numerical differences were observed for several lymphocyte subsets (including CD45⁺, CD3⁺, CD3⁺CD4⁺, CD4⁺CD38⁺, and CD8⁺CD38⁺) between deceased and surviving patients, these differences did not attain statistical significance in our cohort. Kaplan-Meier analysis also revealed no significant difference in OS between the defined risk groups. Several factors may account for these non-significant findings. First, the relatively limited sample size may have reduced the statistical power to detect associations with modest clinical effect sizes. Furthermore, the use of the median value as a uniform cutoff for all biomarkers—a conservative strategy adopted to mitigate overfitting in a small cohort—may have obscured the predictive utility of some markers. Consequently, validation of these results will require future multi-center studies with larger sample sizes or well-designed prospective trials.

In contrast to the non-significant markers, our analysis identified several lymphocyte subsets with clear prognostic value. Patients with higher proportions of CD8⁺CD28⁺ T cells had significantly longer OS, a finding consistent with reports by Liu et al. in non-small cell lung cancer (NSCLC) that affirm the central role of these effector T cells in anti-tumor immunity (41, 42). Furthermore, activated CD4⁺HLA-DR⁺ T cells were significantly enriched in the survival group and independently predicted a better prognosis. The importance of such highly activated T cells has also been validated in studies on NSCLC and in melanoma cohorts receiving anti-PD-1 therapy, further emphasizing the key value of activated T cells in anti-tumor immunity (43, 44). Collectively, these results underscore the prognostic significance of specific activated T-cell populations.

The biological interpretation of complex omics data has long been a primary challenge in clinical translational research (16). The emergence of ML algorithms provides an effective tool for discovering potential biomarkers from these high-dimensional datasets and constructing predictive models (45–47). For instance, a 10-metabolite GC diagnostic model developed by Chen et al. using ML analysis demonstrated performance significantly superior to models based on traditional clinical parameters (16). In OS research, Zhao et al. successfully developed a diagnostic model for identifying high-risk subtypes by integrating transcriptomic and methylation data with ML algorithms, which showed good diagnostic performance (48). To optimize the model and avoid overfitting, this study initially employed seven different algorithms to construct models containing combinations of 1 to 5 variables. The results revealed that algorithms such as GBM and XGBoost achieved AUC values exceeding 0.90 with only two variables, substantially outperforming the traditional AIC algorithm. This highlights the advantage of ML in data dimensionality reduction and building streamlined yet robust prognostic models. The stability of the model was further confirmed through internal Bootstrap validation and 5-fold CV. Ultimately, the risk grouping defined by integrating lymphocyte subsets via ML was confirmed as an independent prognostic factor for OS patients. Its significant clinical value lies in its ability to identify a subgroup of patients without detectable metastasis at initial diagnosis who nonetheless harbor a high risk of mortality.

It is important to acknowledge the limitations of this study. First, the retrospective and single-center nature of the cohort (n=65 high-grade osteosarcoma patients) introduces inherent risks of selection bias and limits the external generalizability of the findings. While osteosarcoma is a rare disease, making large cohorts challenging to assemble, and the institution’s peripheral blood lymphocyte subset analysis was only recently implemented, these factors underscore the need for cautious interpretation. Furthermore, despite employing machine learning algorithms (GBM) and rigorous internal validation strategies (bootstrap and 5-fold cross-validation) to enhance model robustness and mitigate overfitting risks in this modest dataset, the absence of an independent external validation cohort remains a critical constraint. Future validation through larger, multicenter prospective studies is indispensable to confirm the generalizability and clinical utility of the prognostic model. Second, regarding the lymphocyte subset analysis, the panel used in our institution’s flow cytometry did not include Tregs, a crucial subset, preventing the assessment of their role in OS prognosis. Furthermore, with the growing understanding of the TIME, characterizing T-cell function using only dual markers such as CD4+CD38+ or CD8+CD38+ may be insufficient. For instance, studies indicate that CD8+CD38+ programmed cell death protein 1-positive (PD-1+) and PD-1− subpopulations may possess distinct functional properties (49).

Nevertheless, by integrating lymphocyte subset data with ML algorithms, this study effectively revealed the potential value of peripheral blood lymphocyte subsets in prognostic assessment for OS. The findings suggest that functional lymphocyte subsets (e.g., activated or exhausted phenotypes) might reflect a patient’s immune status and disease progression risk more accurately than the simple CD4/CD8 ratio alone. Future research should integrate a broader panel of markers (e.g., PD-1, cytotoxic T-lymphocyte-associated protein 4 [CTLA-4]) to deeply characterize lymphocyte functional subsets and clarify their specific roles in personalized prognosis prediction and treatment strategies for OS.

In conclusion, this research introduces a pioneering framework for OS prognosis by leveraging ML to decode peripheral blood lymphocyte subset signatures. This model addresses the shortcomings of conventional prognostic markers and provides a robust theoretical and practical basis for future investigations into the evolving TIME, ultimately aiming to improve clinical decision-making and patient survival.