Osteosarcoma is the most common primary malignant bone tumor, accounting for 35% of all primary bone tumors, with an incidence that has been rising at approximately 0.3% per year over the past decade. It predominantly affects children and adolescents, typically arising in the metaphyses of long bones in the limbs. Despite advances in multiagent chemotherapy and limb-salvage surgery, the 5-year survival rate remains poor for patients with metastatic or recurrent disease. Adjuvant chemotherapy, while improving survival outcomes, carries significant toxicities, with myelosuppression-related anemia among the most consequential.
The clinical problem: Cancer-related anemia (CRA) is a decline in red blood cell count or hemoglobin attributable to the tumor itself or to radiotherapy and chemotherapy. Approximately 50% of all cancer patients exhibit anemia symptoms, and this figure rises to 90% among those with advanced tumors or undergoing chemotherapy and radiotherapy. In osteosarcoma specifically, post-chemotherapy anemia directly compromises prognosis by impairing oxygen delivery to tissues, reducing treatment tolerance, and creating conditions that favor tumor progression. Early detection and accurate risk stratification for anemia would allow clinicians to intervene proactively with blood transfusions, erythropoiesis-stimulating agents, or dose modifications before hemoglobin levels drop to dangerous thresholds.
The gap this study fills: Despite a growing body of literature applying machine learning to osteosarcoma prognosis, no prior studies had used AI specifically to predict post-chemotherapy anemia risk in this population. This paper, published in Cancer Medicine in 2024 and conducted at Guangxi Medical University Cancer Hospital, addresses that gap by applying five machine learning methods to clinical and laboratory data from 631 osteosarcoma patients, with the goal of building and validating a practical risk prediction tool and interactive nomogram.
The research was supported by the Guangxi Natural Science Foundation Joint Project on Regional High-Incidence Diseases and the Guangxi Zhuang Autonomous Region Health Commission. The final validated model integrates five routinely available laboratory markers into an online nomogram accessible at anemic-prediction-of-osteosarcoma.shinyapps.io/DynNomapp/, enabling clinicians to calculate individualized anemia probability at the point of care.
The study retrospectively collected data from 631 osteosarcoma patients admitted to Guangxi Medical University Cancer Hospital between January 2013 and June 2021. Inclusion criteria required biopsy-confirmed osteosarcoma, completion of chemotherapy, full hematological test results, standard treatment adherence, and complete follow-up data. After applying exclusion criteria (no chemotherapy, presence of concurrent tumors, incomplete laboratory data, or lost follow-up), 314 patients were excluded, leaving 317 eligible patients. Of these, 187 were male and 130 female, with ages ranging from 4 to 75 years and a mean age of 26.1 years, reflecting the characteristic young age distribution of osteosarcoma. The cohort was split into a training set (approximately 70%) and a validation set (approximately 30%) using random division.
Data comprehensiveness: The dataset is notable for its breadth. Researchers collected 24 clinical parameters and 76 laboratory parameters per patient, totaling 100 variables. Clinical parameters included demographic information (gender, age, marital status, nationality), anthropometric measures (blood pressure, height, weight, BMI), medical history (smoking, drinking, prior surgery, primary tumor site, TNM stage, Ki-67 expression), and treatment modalities (chemotherapy, radiotherapy, targeted therapy, immunotherapy). Laboratory parameters spanned complete blood counts, coagulation profiles, liver and kidney function tests, inflammatory markers, electrolytes, tumor markers, immunoglobulins, thyroid function tests, lipid panels, and immune cell subset counts.
Statistical preprocessing: Before applying machine learning, general statistical analysis was performed to identify parameters significantly associated with anemia (t-test for normally distributed data, Mann-Whitney U test for non-normal distributions, chi-square test for categorical data). Of the 100 parameters tested, 25 showed statistically significant differences between anemia and non-anemia groups (p less than 0.05), including Gender, Surgery Type, LYM%, MO%, LYM, BLT, TP, ALB, SF, K, Ca, CREA, UA, PT, APTT, FIB, D-dimer, CRP, hs-CRP, ESR, PCT, IgA, CEA, and CA-153. Correlation analysis was then conducted on these 25 factors to identify and address collinearity before model construction.
Six-model architecture: The study constructed six distinct prediction models in sequence. Model 1 used single-factor logistic regression on 22 significant variables. Model 2 used multifactor logistic regression. Model 3 applied random forest (RF). Model 4 applied support vector machine (SVM). Model 5 applied LASSO regularization regression. Model 6, the intersection model, was built from the five variables that appeared as important predictors across all preceding models, optimizing parsimony and clinical practicality.
Each of the five primary models was evaluated using the area under the receiver operating characteristic curve (AUC), calibration curves, decision curve analysis (DCA), and clinical impact curves (CIC). AUC values above 0.5 indicate predictive capability, with values approaching 1.0 indicating near-perfect discrimination between anemia and non-anemia cases. All five models demonstrated strong performance in the training set, with AUC values ranging from 0.851 to 0.903.
Logistic regression models: The single-factor logistic regression model (Model 1), built on 22 significant variables, achieved an AUC of 0.895 in the training set. The multifactor logistic regression model (Model 2), which refined the variable list to 8 significant predictors through stepwise selection, achieved an AUC of 0.872. The reduction in AUC from single-factor to multifactor reflects the expected trade-off between model complexity and parsimony: multifactor regression removes redundant variables but may lose some discriminatory information carried by correlated predictors.
Random forest: The RF model (Model 3) used 10-fold cross-validation to determine the optimal number of trees and feature selections (mtry parameter). When mtry was set to 3, the misclassification rate was minimized. The two most important features by "%IncMSE" (percentage increase in mean squared error when a variable is permuted) were ALB (23.4) and ESR (15.0), followed by Ca (13.6), CREA (13.3), D-dimer (12.1), UA (12.3), and SF (7.1). The RF model retained these 7 variables and achieved an AUC of 0.882 (95% CI: 0.808-0.892, p less than 0.05).
SVM and LASSO: The SVM model (Model 4) used recursive feature elimination (SVM-RFE) with 10-fold cross-validation, selecting all 25 initial factors as optimal and achieving an AUC of 0.903 (95% CI: 0.869-0.937, p less than 0.05), the highest among all individual models. The LASSO model (Model 5), which applies L1 regularization to penalize and shrink coefficients of less important variables toward zero, identified 19 relevant indicators and achieved an AUC of 0.902 (95% CI: 0.868-0.936, p less than 0.05). In the training set, performance was ranked: SVM (0.903) greater than LASSO (0.902) greater than single-factor LR (0.895) greater than multifactor LR (0.872) greater than RF (0.882).
While individual machine learning models each identified overlapping but distinct sets of predictors, the study's most clinically actionable contribution is the intersection model (Model 6), built from the five variables consistently identified as significant across all five preceding approaches: albumin (ALB), calcium ion (Ca), creatinine (CREA), D-dimer, and erythrocyte sedimentation rate (ESR). The rationale is direct: variables that survive across five different modeling strategies with different assumptions and regularization behaviors are likely to be genuinely important rather than artifacts of any single method's preferences.
Model 6 performance: The intersection model achieved an AUC of 0.850 in the training set. Although this is somewhat lower than the individual models' best performance (0.903 for SVM), the difference was not statistically significant on DCA comparison, and the model's substantially reduced complexity offers major practical advantages. With only 5 predictors, all of which are routine laboratory markers available in essentially any hospital setting, the model eliminates the data collection burden associated with 19-25 variable models.
Validation set performance: In the independent validation set of 93 patients (48 with anemia, 45 without), the intersection model achieved an AUC of 0.802 (95% CI: 0.713-0.891). The calibration curve demonstrated good agreement between predicted probability and actual observed anemia rates, confirming the model generalizes reasonably well to new patients. The DCA showed significant net benefit across threshold probabilities between 30% and 80%, meaning that for clinicians who would consider intervening when the model predicts a 30-80% anemia probability, using the model leads to better decisions than treating all patients or treating none. Below a 30% threshold, the model's benefit was comparable to treating all patients as predicted positive.
Nomogram for individualized risk: A column-chart nomogram was generated from the multivariate logistic regression of the five intersection variables. Clinicians assign each patient a score for each of the five variables, sum the scores, and read the predicted anemia probability from a scale. An online interactive version at anemic-prediction-of-osteosarcoma.shinyapps.io/DynNomapp/ accepts individual patient values and outputs a personalized probability estimate, making the model immediately deployable without specialist biostatistics knowledge.
Albumin (ALB): Albumin is the most abundant plasma protein, with a central role in maintaining oncotic pressure and transporting nutrients and hormones. In the multifactor logistic regression, ALB showed a significant negative association with anemia (OR 0.852, 95% CI 0.748-0.971, p = 0.016), meaning lower albumin is associated with higher anemia risk. The authors propose two mechanisms. First, osteosarcoma as a wasting disease escalates albumin catabolism, producing malnutrition that impairs the erythropoietic microenvironment. Second, diminished albumin may disrupt the transport of iron and essential nutrients to bone marrow progenitors, directly impairing red blood cell production. Across all five ML models, ALB was the most consistently top-ranked predictor.
Calcium ion (Ca): Ca showed a strongly negative association with anemia (OR 0.047, 95% CI 0.004-0.536, p = 0.014), indicating that hypocalcemia is associated with markedly elevated anemia risk. Calcium regulates cellular function through concentration gradients across cell membranes. The authors hypothesize that post-chemotherapy osteosarcoma patients may experience disruptions to calcium homeostasis, potentially through bone resorption effects, which alter cellular signaling environments relevant to hematopoiesis.
Creatinine (CREA): Elevated CREA was associated with higher anemia risk (OR 0.974, 95% CI 0.948-1.000, p = 0.050). Chemotherapy nephrotoxicity is well-established, and renal injury reduces secretion of endogenous erythropoietin, the primary hormonal driver of red blood cell production. Reduced erythropoietin shortens red blood cell lifespan, impairs bone marrow hematopoietic function, and reduces circulating red blood cell mass, directly causing anemia. Elevated serum creatinine serves as a proxy for this renal-erythropoietic axis disruption.
D-dimer and ESR: D-dimer (OR 1.188, 95% CI 1.052-1.341, p = 0.005) and ESR (OR 1.031, 95% CI 1.010-1.052, p = 0.003) both showed positive associations with anemia, meaning elevated levels predict higher risk. D-dimer reflects concurrent activation of coagulation and fibrinolysis, a state characteristic of malignancy-associated hypercoagulability. Osteosarcoma cells can directly activate prothrombin, and rapid tumor growth damages surrounding vasculature, perpetuating a cycle of clotting and fibrinolysis that consumes red blood cells and platelets. ESR reflects the sedimentation rate of erythrocytes, which accelerates in anemia (fewer cells, less plasma resistance) and with elevated inflammatory cytokines released during post-chemotherapy immune compromise. High ESR is therefore both a consequence of and a contributor to the inflammatory milieu that suppresses erythropoiesis.
While the intersection model uses only five biomarkers, the full multifactor logistic regression model (Model 2) identified nine independently significant predictors after controlling for confounders. Beyond the five intersection variables, four additional factors emerged: lymphocyte count (LYM, OR 0.451, 95% CI 0.221-0.918, p = 0.028), platelet count (BLT, OR 1.005, 95% CI 1.001-1.009, p = 0.016), serum ferritin (SF, OR 0.852, 95% CI 0.748-0.971, p = 0.016), and procalcitonin (PCT, OR 1.667, 95% CI 1.114-2.493, p = 0.013). Cancer antigen 153 (CA-153) also reached significance (OR 1.065, 95% CI 1.004-1.129, p = 0.037).
Lymphocyte count: LYM showed a strong protective association (OR 0.451), meaning lower lymphocyte counts substantially increase anemia risk. This is biologically coherent: chemotherapy causes widespread lymphodepletion, and severe lymphopenia reflects profound immune suppression that correlates with the depth of bone marrow toxicity. Patients with more severe myelosuppression are more likely to have concurrent anemia as part of pancytopenia.
Platelet count: BLT showed a positive association (OR 1.005, p = 0.016). This appears counterintuitive at first, but elevated platelet counts in cancer patients often reflect reactive thrombocytosis driven by chronic inflammation and tumor-derived thrombopoietic factors. A higher baseline platelet count may proxy for higher tumor burden and more active inflammatory signaling, both of which independently drive anemia through myelosuppression and cytokine-mediated erythropoietic inhibition.
Procalcitonin (PCT): PCT (OR 1.667, p = 0.013) is traditionally a marker of bacterial sepsis, but elevated PCT in cancer patients can also reflect tumor necrosis and severe immunosuppression. Higher PCT in post-chemotherapy osteosarcoma patients likely signals more severe infectious complications and inflammatory burden, conditions that suppress bone marrow function and accelerate red blood cell destruction. Serum ferritin (SF) showed a weak positive association, consistent with its role as an acute-phase reactant elevated in inflammation and as an indicator of iron dysregulation in the bone marrow.
Single-center retrospective design: All 317 included patients were drawn from a single institution, Guangxi Medical University Cancer Hospital, over an eight-year period. Single-center cohorts may reflect institutional practices in chemotherapy protocols, laboratory measurement platforms, and patient demographic characteristics that differ substantially from centers in other regions or countries. The validation set was created by random splitting within this same institutional cohort rather than drawn from an independent external center, meaning both training and validation sets share the same underlying population distribution. This internal validation approach consistently overestimates real-world generalizability compared to external validation across geographically or clinically distinct institutions.
Sample size limitations: After exclusions, only 317 patients were included, with 146 in the anemia group and 171 in the non-anemia group in the training set, and 48 anemia versus 45 non-anemia in the validation set. While this imbalance is modest, machine learning models applied to relatively small datasets are vulnerable to overfitting, particularly for algorithms like SVM and LASSO that operate on high-dimensional input spaces. The gap between SVM's training AUC (0.903) and the intersection model's validation AUC (0.802) illustrates how model performance can degrade when tested on truly new data.
Absence of prospective validation: All data were collected retrospectively, meaning that clinicians were not using the model in real time and patient management was not influenced by its predictions. Prospective validation, in which the model is embedded into clinical workflows and applied to new patients as they present, is needed to assess whether it would actually improve clinical decision-making and outcomes, rather than simply achieving favorable statistical metrics on historical data.
Correlation-driven redundancy: The pre-model correlation analysis identified significant pairwise correlations among several predictors (for example, surgery type and surgery performance at r = 0.90; LYM and LYM% at r = 0.52; PT and APTT at r = 0.53), and the study excluded some redundant variables accordingly. However, the multifactor logistic regression and other models may still incorporate partially correlated variables, which can inflate standard errors, reduce coefficient stability, and complicate interpretation of individual predictor effects.
Multicenter external validation: The authors explicitly call for prospective multicenter investigations to expand the clinical applicability and robustness of the intersection model. Recruiting patients from hospitals with different ethnic compositions, chemotherapy regimens, and laboratory platforms would test whether the five-variable nomogram maintains acceptable discrimination (AUC above 0.75) across diverse populations. Given that the study population was predominantly drawn from Guangxi Zhuang Autonomous Region, where ethnic minorities including Zhuang, Yao, Miao, Mulao, and Han nationalities were represented, external validation should include populations with different demographic distributions to assess model stability.
Dynamic and longitudinal monitoring: The current model predicts anemia risk based on pre-chemotherapy or baseline laboratory values. A natural extension would be serial monitoring, in which the model is re-applied at multiple chemotherapy cycle timepoints to generate dynamic risk trajectories. Changes in ALB, Ca, CREA, D-dimer, and ESR across cycles could potentially identify patients whose risk is escalating in time to intervene before severe anemia develops, rather than predicting anemia risk only at a single baseline assessment.
Integration with treatment planning: The five-variable nomogram could be incorporated into clinical decision support systems embedded in electronic health records, generating automated anemia risk alerts when laboratory results are entered. For high-risk patients identified before each chemotherapy cycle, clinicians could preemptively order hemoglobin monitoring at shorter intervals, schedule earlier transfusion or erythropoiesis-stimulating agent consultations, or consider dose modification protocols. This kind of proactive management is particularly valuable because post-chemotherapy anemia in osteosarcoma directly impairs patient performance status and tolerance for subsequent chemotherapy cycles.
Expanding the feature set: Future iterations could incorporate genomic and transcriptomic markers, imaging features from staging CT or MRI, and pharmacokinetic data from therapeutic drug monitoring. Given the strong predictive importance of ALB across all five models, nutritional interventions (albumin infusions, high-protein supplementation) pre-chemotherapy could be tested as prophylactic strategies in patients identified as high-risk, with anemia incidence as the primary endpoint in prospective trials. The publicly accessible online nomogram provides a ready-made platform for such clinical trial enrichment strategies.