Baseline parameters and potential diagnostic biomarkers
All 323 surgical patients were analyzed in our study. Among these patients, 131 patients were eventually excluded due to incomplete baseline data and Intracranial multiple lesions. Therefore, a total of 192 patients were entered in our cohort including 70 patients with GBM, 41 patients with PCNSL and 81 patients with BM.
We collected 70 parameters from the routine preoperative blood examination and demographic characteristics, of which 23 parameters were found have significant differences among GBM, sBM and PCNSL. These parameters were listed on table 1 and table2. The median age (IQR, interquartile range) for patients with GBM, PCNSL and sBM was 53 years (41-64years), 59.5 years (48-64years) and 59.5years (52-67years), respectively. The significant difference in patient age(p=0.015) was observed between GBM, PCNSL and sBM group. The statistically significant difference was also found in blood routine test parameters such as white blood cells (WBC), Neutrophil value and so on (Table1). The same phenomenon was found in biochemistry parameters such as lactate dehydrogenase (LDH), β2-macroglobulin (β2-MG) and so forth (Table1).
Calculated laboratory parameters like NLR (p=0.048) and coagulation function parameters like, fibrin degradation product (FDP) (p=0.015), TT (p<0.001), INR (p<0.001) were different between the three brain tumors (Table1, Table2). However, 23 predict parameters were inconvenient for differential diagnosis, a more efficient and laconic method should be adopted.
Filtration of Diagnostic biomarkers
In order to find a more efficient way, we used a multinomial logistic regression model as a filtration tool to choose the most important diagnostic predictors among the 23 parameters. Further investigation was made by the filtration model, then we found that the diagnosis model accuracy has a negative correlation with the cut off value. When the cut off value decreased, the predict parameter number and the accuracy of diagnosis increased. As is shown on Figure 2, the accuracy increase trend starts to slow down when the number of parameters is more than 8 and cut off value is below 0.48. Therefore, we chose the 8 parameters as predict factors into multinomial logistic regression. The 8 parameters were patient age, PCT, LDH, β2-MG, α2- globulin (α2-G), INR, TT and FDP. These parameters were all different between GBM, PCNSL and brain metastases (Figure1).
Multinomial logistic regression analysis of diagnostic factors
Age, PCT, LDH, β2-MG, α2-G, INR, TT and FDP were put into multinomial logistic regression analysis to find useful diagnostic factors for differentiating GBM, PCNSL and sBM. In multinomial logistic regression, we set GBM as the reference object. After calculation, the outcome was presented in Table 3. We found patient age, PCT, INR and TT were each significantly associated with differentiating diagnosis of the three brain tumors. In differentiating GBM, PCNSL and sBM, we determined that patient age was independently associated with sBM (OR=1.055, 95%CI 1.016-1.094, p=0.005), whereas PCT was independently negative associated with sBM (OR=0.008, 95%CI 0.004-0.017, p=0.027). Furthermore, while classifying the INR and TT, we found that INR has an independently positive correlation with PCNSL and sBM when compared with GBM, however, TT has a negative correlation with PCNSL and metastases. Compared with patients with level 2 (≥0.99) INR, level 0 (OR=0.11, 95%CI 0.029-0.413, p=0.001) and level 1 (OR=0.119, 95%CI 0.032-0.442, p=0.001) patients were more prone to be diagnosed as GBM than sBM. In contrast, compared with patients with level 2 (≥19s) TT, level 0 (OR=16.31, 95%CI 5.188-51.26, p<0.001) and level 1 (OR=6.455, 95%CI 2.07-20.13, p=0.001) patients were more prone to be sBM than GBM. The same trend was observed when PCNSL compared with GBM, patients with INR level 0 (OR=0.166, 95%CI 0.044-0.619, p=0.007) and level 1 (OR=0.181, 95%CI 0.047-0.694, p=0.013) were more prone to be GBM than patients with INR level 2. And patients with TT level 1 (OR=3.566, 95%CI 1.231-10.331, p=0.019) were more prone to be PCNSL than patients with TT level 2. The other parameters in this analysis had statistically insignificant relationships with either PCNSL or sBM.
Evaluation of efficacy of the preoperative diagnostic model
The formula for diagnostic model was: PrY1 = In (PCNSL/GBM) = intercept + , PrY2 = In (sBM/GBM) = intercept + , PrY3 = 0 (reference). “intercept” represents constant of the regression model, “b0…bn” represent the coefficients of the regression model and x1…xn represent the predictor variables. All these parameters were listed in table 3. We put the 8 parameters into this formula to calculate PrY1, PrY2 and PrY3 value. Then we used the formula: P1=exp(PrY1)/[exp(PrY1)exp(PrY2)exp(PrY3)],P2=exp(PrY2)/[exp(PrY1) exp(PrY2) exp(PrY3)], P3=exp(PrY3)/ [exp(PrY1) exp(PrY2) exp(PrY3)] to calculate possibility of PCNSL (P1), sBM (P2) and GBM (P3). The diagnostic model chose the maximum value as the result.
The accuracy rate of diagnosis for this multinomial regression model was presented by Table 4. Diagnosis accuracy was best in sBM (correct percent=88.2%). GBM diagnosis accuracy rate was 76.1%. However, diagnosis accuracy of PCNSL was lowest, only 22%. This multinomial logistic regression may be more useful in differentiating GBM and sBM.