Background
Bone cancer called osteosarcoma (OS), especially its fibroblastic type, makes things very hard in the world of bone diseases. This happens because of its fierce character and the complexity involved in deciding outcomes. Current prognostic models, like the American Joint Committee on Cancer (AJCC) system and Tumor Node Metastasis (TNM) Staging System, don't always fully include important individual patient factors such as age, sex and race. These things are very important for making a correct prediction.
Methods
A total of 394 patients with fibroblastic osteosarcoma were included in the study, adhering to specified inclusion and exclusion criteria. The cohort was subsequently segregated into training and validation sets at a 7:3 ratio. X-tile software facilitated the determination of optimal age and tumor size cutoffs. Missing data were managed using multiple imputation and K-Nearest Neighbor (KNN) methods. The primary endpoint was cancer-specific survival (CSS), categorized into binary data (survival status at 3 and 5 years) and time-to-event data. Independent prognostic factors were ascertained using the Boruta algorithm, which informed the construction of predictive models employing Cox regression and diverse machine learning algorithms such as Survival Tree, Extra Survival Trees, Random Survival Forest, Gradient Boosting Survival Analysis, Fast Kernel Survival SVM, and Minlip Survival Analysis. Model performance metrics included the concordance index (C-index), accuracy, recall, F1 score, and time-dependent Area Under the Curve (AUC). A calibration plot was generated to validate the accuracy of the most proficient machine learning model. Decision curve analysis (DCA) was implemented to ascertain the model's clinical utility. Additionally, we used the SHapley Additive exPlanations (SHAP) method to show how important our model found key things that can predict outcomes.
Results
For age, the determined optimal cutoff points were established at 40 and 57 years. Regarding tumor size, these points were set at 60mm and 103mm. Our study identified nine significant independent prognostic factors impacting the cancer-specific survival in patients with fibroblastic osteosarcoma. These included age group, tumor stage, tumor size group, radiation, surgery type, primary site, sex, chemotherapy, and grade group. Comparative analysis of different algorithms, utilizing metrics such as accuracy, recall, F1 score, C-index, and time-dependent AUC, highlighted the Extra Survival Trees model as the superior predictive tool for machine learning. This model demonstrated high efficiency (3-year CSS accuracy: 0.91, 5-year CSS accuracy: 0.89), notable recall rates (3-year: 0.81, 5-year: 0.74), and robust F1 scores (3-year: 0.83, 5-year: 0.80), along with an average AUC of 0.89 and a C-index of 0.92 for training and 0.80 for validation. The calibration curve for this model indicated high predictive accuracy, and its clinical usefulness was further corroborated by decision curve analysis (DCA). SHAP analysis identified 'age group', 'tumor stage', and 'tumor size group' as the three most influential variables impacting cancer-specific survival predictions in fibroblastic osteosarcoma. Our study suggested otherwise than previous ones. It showed that radiation and chemotherapy may not work for treating this type of bone cancer called fibroblastic osteosarcoma.
Conclusion
Research indicates that predictive analysis using machine learning outperforms traditional methods in forecasting outcomes for patients with fibroblastic osteosarcoma. This development offers considerable promise for enhancing tailored therapeutic approaches and prognostic outcomes in fibroblastic osteosarcoma.
