A Novel Prognostic Model to Predict Prognosis of Patients With Osteosarcoma Based on Clinical Characteristics and Blood Biomarkers

Background: To develop a novel prognostic model based on clinical characteristics and blood biomarkers to estimate overall survival (OS) and progression-free survival (PFS) in osteosarcoma (OSC) patients. Methods: A total of 71 patients with OSC from Sun Yat-sen University Cancer Center were retrospectively included. The novel prognostic model for predicting OS and PFS was established by using Lasso regression analysis based on blood biomarkers. The predictive accuracy and discriminative ability of the novel prognostic model was compared with TNM staging and clinical treatment using concordance index (C-index), time-dependent ROC (tdROC) curve, decision curve analysis (DCA), net reclassication improvement index (NRI), and integrated discrimination improvement index (IDI). Results: Based on the Lasso regression analysis, we identied a 5 prognostic factors module (RBC, Ca 2+ , CRE, PNI, and LSR) as a novel predictive model for the OSC patients. The C-index of the novel prognostic model for predicting OS and PFS were 0.782 (95% CI = 0.658 - 0.905) and 0.741 (95% CI = 0.632 - 0.851), respectively, which were higher than that of TNM staging and clinical treatment. The tdROC curve and DCA also showed the novel model had good predictive accuracy and discriminatory power than TNM staging and treatment both in predicting OS and PFS. Moreover, the novel prognostic model performed well in all time frames (1, 3 and 5 years) in terms of the IDI and NRI when compared with the TNM staging, and clinical treatment. Conclusions: The novel prognostic model showed favorable performance than TNM staging and clinical treatment for predicting OS and PFS in OSC patients. ratio; APOA: apolipoprotein A1; APOB: apolipoprotein B; ABR: APOA / APOB ratio; AI: The atherogenic index; GLU: glucose. treatment, and scheduling patient follow-up. PNI:

shown to be positively correlated with tumor volume, which has additional useful prognostic signi cance [13]. So far, however, few researchers used combination of clinical characteristics and blood biomarker to predict the prognosis of OSC.
Thus, the aim of the present retrospective study was to develop a novel prognostic model based on clinical characteristics and blood biomarkers to estimate overall survival (OS) and progression-free survival (PFS) in patients with OSC and to assess its incremental value in traditional staging systems and clinical treatment of individual OS and PFS.

Patients Follow Up
The patients' survival data follow-up was obtained by means of retrieving medical records, email, and direct telecommunication, all patients were followed up until death or July 2020, if still alive. Overall survival (OS) was de ned as the time interval from diagnosis to the date of the patient's death or censored at the date of last follow-up. Progression-free survival (PFS) was calculated from the date of diagnosis to the date of the objective disease progression or death or the date of the last follow-up visit.

Statistical analysis
Statistical analysis was performed using R software (version 3.6.1, https://www.r-project.org). Continuous variables were presented as mean ± SD and tested by t-test or Wilcoxon test. Firstly, we utilized the LASSO regression algorithm (λ were determined by 10-fold cross validation with the error of the minimum criteria) to select the most useful prognostic factors out of all the OSC-associated blood biomarkers, and then established a novel predictive model. Subsequently, the predictive accuracy and discriminative ability of the novel prognostic model was compared with TNM staging and clinical treatment using Harrell concordance index (C-index), time-dependent ROC (tdROC) curves [18], decision curve analysis (DCA) [19], net reclassi cation improvement index (NRI), and integrated discrimination improvement index (IDI) [20]. The larger C-index and area under the curve (AUC) of tdROC curves, the better the model is for the risk prediction. DCA was used to evaluate the clinical usefulness and net bene t of the predictive model [21]. The NRI assessed the ability of a new model to re-classify subjects compared to an old model into binary event or no-event categories. The IDI index quanti ed the improvement in average sensitivity without reducing the average speci city of a new model compared with an older model [22]. The correlation between the novel prognostic model, TNM staging and clinical treatment was evaluated by Pearson's correlation coe cient. Besides, a nomogram integrating the prognostic model risk score, TNM staging, and clinical treatment was developed that may assist in individual survival prediction of OSC patients. Internal validation and calibration of the nomogram were conducted by 1000-resample bootstrapping. Finally, we illustrated the discrimination by dividing the patients into low-risk groups and high-risk groups according to the novel predictive model scores. The Kaplan-Meier method was used to perform OS and PFS analysis, and the log-rank test was used to compare signi cance of the differences of survival distribution between groups. Generally, a P value of ≤ 0.05 was considered as statistically signi cant for all analyses.

Construction Of Prognostic Model For Os And Pfs
Firstly, the LASSO regression analysis was performed to extract signi cant predictors associated with overall survival (OS). Figure 1A showed the change in trajectory of each predictor was analyzed. Afterwards, the optimal value for λ was determined using 10-fold cross-validation with the minimum criteria (Fig. 1B). According to the criteria, the optimal value of the λ was 0.076 in this study, and its corresponding predictors were considered to be the signi cant prognostic factors for OS, which included RBC, Ca 2+ , CRE, PNI, and LSR. Finally, a prognostic model was constructed for predicting OS and PFS based on the coe cients of signi cant predictors derived from the LASSO regression, with a risk score was calculated by Firstly, we calculated the C-index in the three predictive signatures, as shown in Table 2. For OS, the C-index of the novel prognostic model was 0.782 (95% CI = 0.658-0.905), which was higher than that of the TNM staging  (Fig. 2). Thirdly, the DCA showed the novel prognostic model had a higher overall net bene t than TNM staging and clinical treatment across the majority of the range of reasonable threshold probabilities both in OS and PFS (Fig. 3). Finally, both the NRI and IDI calculations were obtained at 1, 3 and 5 years and used to compare the alternative prognostic indices of our model with the TNM staging and clinical treatment. The results were presented in Table 3. For OS, NRI analysis revealed that the accuracy of the novel prognostic model was higher than that of the TNM In addition, the similar results also showed that the novel prognostic model had a good performance in predicting the PFS for OSC patients than others.  The nomogram incorporating the prognostic model risk score, TNM staging, and clinical treatment to predict the probability of 1-, 3-, and 5-year OS (Fig. 4A) and PFS (Fig. 4B) in OSC patients. Each patient would assign one point for each prognostic variable, the estimated probability of 1-, 3-and 5-year OS and PFS was determined by summing all of the point, and the higher number of total points indicated a worse outcome for the patient. In addition, the calibration curve showed good agreement between prediction and observation in 1-, 3-, and 5-year OS (Fig. 4C, 4E, 4H) and PFS (Fig. 4D, 4F, 4I). The Survival analyses of OSC patients according to prognostic model risk score Using the R package "survminer" and "survival", we classi ed patients into low -risk patients and high-risk patients based on the prognostic model risk score, and make the Kaplan-Meier curve. The results showed that patients with higher risk scores (risk score > -0.94) had a signi cantly lower OS (Fig. 6A, P < 0.001) and PFS (Fig. 6B, P = 0.001) rate than their lowrisk counter-parts (risk score ≤ -0.94). In order to test whether the prognostic model could remedy the current de ciencies of AJCC TNM stage. Patients were factitiously strati ed into early stage (stage I/II) and late stage (stage III/IV). Kaplan-Meier curve showed that high-risk patients in the early stage had signi cantly shorter OS (Fig. 6C, P < 0.001) and PFS (Fig. 6D, P < 0.001) than low-risk patients, but in the late stage, the OS (Fig. 6E, P = 0.220) and PFS (Fig. 6F, P = 0.450) in low-risk patients and high-risk patients displayed no signi cant difference.
The serum levels for the 5 selected predictors in the low-risk and high-risk patients

Discussion
In the present study, we had analyzed individual clinical characteristics and blood biomarkers based on survival analysis.
The Lasso regression algorithm was used to establish a novel prognostic model for predicting OS and PFS in OSC patients. Compared with traditional TNM staging and clinical treatment, our prognostic model had better predictive accuracy and discriminatory ability. The prognostic model successfully classi ed those patients into high-risk and low-risk subgroups that were signi cantly different in terms of OS and PFS.
According to the results of LASSO regression analysis, we identi ed a 5 prognostic factors module (RBC, Ca 2+ , CRE, PNI, and LSR) as a novel predictive model for the OSC patients. RBC count was one of the erythrocyte parameters, the reduced preoperative RBC count might re ect worse liver function, it was well known that liver function impacts patients' survival [23,24]. Lu et al. had reported that preoperative RBC counts lower than normal had worse OS rates than those without reduced preoperative RBC counts in primary liver cancer patients [25]. Ca 2+ played a pivotal role in cancer cells growth, migration, and death, serving as a principal signalling agent and the expression of Ca 2+ channel transcripts had been highlighted as a potential biomarker in the growing number of cancers [26,27]. CRE as a marker of kidney function, which had been investigated as a prognostic parameter in colorectal [28], liposarcoma [29], and prostate cancer [30]. PNI could re ect the immune and nutritional status of human body, previous studies had suggested the predictive and prognostic value of it in a variety of tumors [31,32]. Huang et al. found that preoperative low PNI was signi cantly correlated with OSC tumor size, tumor staging, pathologic fractures, local recurrence, and metastasis, suggesting that PNI may be an important prognostic parameter in patients with OSC [33]. LSR was often used in the assessment of liver injury, the increased serum LSR in patients with gastric adenocarcinoma was associated with better prognosis [34]. Chen et al. found that alginate oligosaccharide treatment reduced the progression of OSC, and decreased levels of IL-1, IL-6, and the ratios of AST/ALT, which may be related to the improvement of antioxidant and anti-in ammatory capabilities in OSC patients [35]. All of the above research showed that the 5 prognostic factors were closely related to the occurrence or development of tumors. These suggested that our analysis results had credible prognostic value.
In order to determine whether our model could remedy the de ciencies of TNM staging in the prognostic assessment of OSC patients, we divided the patients into low-risk and high-risk subgroups based on the prognostic model risk scores. Kaplan-Meier survival curves showed that the high-risk groups were associated with shortened survival in OSC patients with stage I&II and stage III&IV. Thus, the results reminded us that even patients in the same stage, high-risk patients required more intensive treatment. Moreover, the results also suggested that our model could remedy the de ciencies of TNM staging and enhance the predictive power of TNM stage. Improved prediction of individual prognosis would help clinicians in counseling patients, selecting personalized treatment, and scheduling patient follow-up.
Compared to previous studies [33,36,37], this study had the following advantages: 1. The prognostic factors of OSC in the past were mostly single indicators, but our study included more potential prognostic factors than previous studies. The simplicity, cheapness and availability of the ve prognostic biomarkers fully re ect the advantages of their combined application. 2. We used the new algorithm LASSO regression analysis to develop a prognostic model as a statistical method for ltering variables to establish a prognostic model, which allows to adjust the over tting of the model to avoid extreme predictions, so the prediction accuracy could be signi cantly improved, and this method has been applied in many studies [38][39][40]. 3. In this study, we used multiple methods to compare the predictive accuracy and discriminative ability of the novel prognostic model with TNM staging and clinical treatment. And these results all showed that our model outperformed than others. 4. The endpoint of this study were OS and PFS, so this model could achieve better clinical application.
However, there still presented some limitations in this study. 1. This was a retrospective analysis, its selection bias may be unavoidable, so its calculated predictive value was for clinicians' reference only. 2. We only analyzed the data from a single cancer center, with a small sample size. Therefore, it is necessary to conduct multi-center and large-scale studies in the future to further verify the generalizability of the prognostic model established in this study. 3. This study had shown that RBC, Ca 2+ , CRE, PNI and LSR were related to the prognosis of OSC patients, but the molecular mechanisms behind the above-mentioned effect have not been clari ed. 4. Although these ve predictors were easy to obtain, it was undeniable that these predictors were all non-speci c OSC predictors and may lack certain speci city. Some biomarkers may be incorporated into these prognostic models to improve speci city, such as immunohistochemical markers [41], radiomics [42,43], and recently newly applied non-coding RNAs [44,45]. 5. We collected data only for the initial diagnosis and did not dynamically monitor the entire course of the patient, so we could not know the signi cance of biomarkers for the prognosis of the patient after each treatment. Despite the above shortcomings, the prognostic model was effective and may help predict the prognosis of OSC patients, providing clinicians with a more practical and convenient tool for individualized treatment decision making and survival assessment at the initial diagnosis.

Conclusions
In conclusion, we successfully established a novel prognostic model based on clinical characteristics and blood biomarkers, which showed outperform TNM staging and clinical treatment in predicting OS and PFS in OSC patients. The