Development and validation of a novel risk model in newly diagnosed de novo bone metastatic prostate cancer (M1b): a retrospective study

Patient information

This was a retrospective, single-institution study approved by the ethical committee of the Affiliated Hospital of Xuzhou Medical University (XYFY2022-KL192-01). After review by the ethics committee, it was agreed to conduct the study according to the reviewed study protocol and the informed consent was waived. Patients who were newly diagnosed with de novo bone metastatic prostate cancer at our institution from January 2008 to June 2021 were enrolled. The inclusion criteria were: (1) pathologically confirmed prostate adenocarcinoma; (2) complete imaging data and imaging-confirmed bone metastases; (3) no prostate cancer-related treatment prior to pathological diagnosis; and (4) treated with ADT or ADT-based therapy, such as surgical/pharmacological castration+anti-androgen therapy (bicalutamide; flutamide)/androgen biosynthesis inhibitor (abiraterone). The exclusion criteria were: (1) visceral metastasis; (2) adenocarcinoma with neuroendocrine differentiation; (3) the presence of other cancers that affect survival; (4) subsequent tumor-reductive surgical treatment; and (5) incomplete clinicopathologic and follow-up data.

The patients’ demographic and clinicopathological data, such as age, height, weight, hypertension (Hp), diabetes mellitus (DM), prostate-specific antigen (PSA), PSA nadir (nPSA), time to PSA nadir (TTN), alkaline phosphatase (ALP), hemoglobin (Hb), fibrinogen (Fib), clinical T stage [cT], and Gleason score were obtained from their medical records. The tumor-node-metastasis (TNM) stage was assigned according to the recommendation (Buyyounouski et al., 2017). Eventually, a total of 122 patients were included in the study.

Bone metastatic evaluation

Bone scan images were obtained before therapy. We re-reviewed all bone scans and recorded the metastases sites and numbers. Previous bone-based risk stratification models, i.e., EOD, EBM and H/LVD, have been described previously (Soloway et al., 1988; Gravis et al., 2016; Eisenberger et al., 1998).

Our model was based on a previous classification of bone metastatic sites (Suzuki et al., 2021) and classified metastasis into six sites: skull; cervical, thoracic, and lumbar vertebrae; chest (ribs and sternum); pelvis; upper limbs; and lower limbs. We calculated the bone metastatic score (BMS) for each site as follows: 0 points were assigned for non-metastasis and 1 point was assigned for metastasis. The scores were summed to obtain the BMS (range: 0–6 points). To group patients according to BMS, we used the X-tile software (version 3.6.1) (Camp, Dolled-Filhart & Rimm, 2004) to obtain optimum cutoff values (High-risk group [HRG]: BMS ≥3; Low risk group [LRG]: BMS < 3).

Follow-up

The patients were followed up mainly by telephone, combined with outpatient and/or inpatient visits. The contents of the follow-up visits were recorded on designated forms. The endpoint was CRPC. The duration was calculated from the date of prostate biopsy. CRPC was defined based on the recommendation (Cornford et al., 2017). The last follow-up date was March 4, 2022 and patients with mHSPC at the last follow-up were censored.

Statistical methods

The numerical variables and categorical variables were described as medians (interquartile ranges [IQRs]) and numbers (percentages), respectively. The TTCRPC was estimated using the Kaplan–Meier method. Numerical variables were converted into binary variables according to reference range of the normal upper or lower limit and/or previous article classification standards (Lin et al., 2019; Kanesaka et al., 2021; Tian et al., 2020). The Cox proportional hazards regression model was used in the univariable and multivariable analyses to calculate the hazard ratios (HRs). Clinicopathological variables that demonstrated a univariable relationship with the TTCRPC (p < 0.100) and were considered clinically relevant were subsequently entered into the multivariable model.

The new bone risk stratification was validated by calculating the area under the receiver operating characteristic (ROC) curve (AUC) using R software to compare the validities of the H/LRG and the other models  (Bouassida et al., 2021). The nomogram for predicting the 1-, 2-, and 3-year CRPC was developed using Cox-derived coefficients. Model calibration and discrimination were assessed with calibration plots and the C-index. The model overfitting was quantified using bootstrapped resampling (1,000).

A two-sided p < 0.050 was considered statistically significant. Analyses were performed using Statistical Product and Service Solutions (SPSS, version 26.0; SPSS, Inc., Chicago, IL, USA) and R (version 4.1.1; R Core Team, 2021).

Baseline characteristics of the patients

Table 1 lists the study population’s demographic and clinicopathological characteristics. The median patient age was 73 years. Forty-two patients (34.4%) presented with hypertension and/or diabetes. Most patients had Gleason score ≤ 8 (93 cases, 76.2%) and high-volume disease (84 cases, 68.9%). The follow-up period was 2–92 months with a median of 11.5 months. Eighty-six patients progressed to CRPC during the follow-up.

Establishment of bone risk stratification model

The most common bone metastatic site was the pelvis (90.2%), followed by the cervical, thoracic, or lumbar vertebrae (69.7%), chest (68.0%), upper limbs (56.6%), lower limbs (49.2%) and skull (28.7%) (Table 2). Univariable Cox analysis revealed that almost every metastatic site (except the pelvis) was an important factor affecting TTCRPC (p < 0.05) (Table 2). The median BMS was 4 (Table 3), where approximately 50% of the patients had scores of ≥4. According to the H/LRG criteria, 76 patients had HRG, while 46 had LRG.

Figure 1 shows HRG had a signifcantly poorer TTCRPC than the LRG(p < 0.001). Figure 2 displays the comparison of our model with the other three risk stratification models (EOD, EBM, and H/LVD) according to their CRPC AUCs. The 1-, 2-, and 3-year AUCs for the EOD, H/LVD, and EBM were 0.573, 0.715, and 0.671; 0.587, 0.710, and 0.829; and 0.596, 0.702, and 0.695, respectively. The 1-, 2-, and 3-year AUCs for our model were 0.620, 0.754, and 0.793, respectively. H/LRG demonstrated higher validity than the other risk stratification models.

Prognostic factors for CRPC

In the univariable analysis, nPSA (>0.2 ng/ml vs. ≤0.2 ng/ml; HR 3.778, 95% CI [2.154–6.627], p < 0.001), H/LRG (HRG vs. LRG; HR 2.644, 95% CI [1.623–4.307], p < 0.001), TTN (≤6 months vs. >6 months; HR 3.158, 95% CI [1.888–5.281], p < 0.001), ALP (>128 U/L vs. ≤128 U/L; HR 2.335, 95% CI [1.512–3.605], p < 0.001), and Gleason score (>8 vs. 8; HR 1.676, 95% CI [1.044–2.689], p = 0.032) demonstrated statistically significant effects on TTCRPC (Table 4).

Considering that ALP demonstrated collinearity with bone metastases, ALP was not included in the multivariable analysis. Finally, five variables (Gleason score, H/LRG, nPSA, TTN, and Fib) were included in a multivariable Cox proportional hazard regression model for the analysis. The results demonstrated that Gleason score (>8 vs. ≤8; HR 1.860, 95% CI [1.125–3.075], p = 0.016), nPSA (≤0.2 ng/ml vs. >0.2 ng/ml; HR 2.384, 95% CI [1.242–4.573], p = 0.009), TTN (≤6 months vs. >6 months; HR 2.243, 95% CI [1.257–4.003], p = 0.006), and H/LRG (HRG vs. LRG; HR 1.864, 95% CI [1.108–3.135], p = 0.019) were independent factors influencing the TTCRPC (Table 4).

Development and validation of nomogram for predicting CRPC

We developed a nomogram based on four parameters (Gleason score, H/LRG, nPSA, and TTN) to predict the 1-, 2-, and 3-year CRPC (Fig. 3). The effectiveness of the model was examined with discrimination evaluation (C-index) and conformity assessment (calibration curve plotting). The C-index was calibrated during the internal validation in R using the bootstrap method, and the 1-, 2-, and 3-year calibration curves were plotted. With the bootstrap = 1,000 setting, the evaluation results demonstrated that the CRPC nomogram prediction model had a C-index of 0.727, indicating that the CRPC nomogram prediction model had good predictive accuracy. Accordingly, the calibration curve of the model was plotted (Fig. 3), and demonstrated that the model had good conformity.