Background

For patients with cancer, the identification and application of prognostic markers can assist in predicting clinical outcomes, facilitate treatment choice, and improve therapeutic research [1]. Various models can be used to predict the risk of the disease using data obtained from cohort studies or from randomized trials in multivariate analyses [2]. The use of machine learning algorithms is expected to improve the definitive identification of prognostic factors, due to its increased flexibility and enhanced performance compared with traditional statistical modeling techniques [3], although there remains room for improvement in machine learning methodology [4].

Castration-resistant prostate cancer (CRPC) is a form of prostate cancer that progresses despite the use of androgen depletion therapy. However, there are limitations associated with the biomarkers currently available for CRPC, adding to the challenges faced by physicians when making prognostic and therapeutic decisions [5]. Patients with CRPC may present with few symptoms but have rising levels of serum prostate-specific antigen (PSA), or they may have multiple metastases and significant morbidity [6]. Between 10 and 20% of patients with prostate cancer develop CRPC within 5 years, and a pooled survival estimate suggested that patients with CRPC could expect to live for 14 months following diagnosis (range 9 to 30 months) [7].

The standard first-line chemotherapy treatment for CRPC recommended by current guidelines includes systemic docetaxel plus concurrent steroids [6, 8, 9]. Data from clinical trials demonstrated benefits for enzalutamide (a novel androgen receptor signal inhibitor) [10], abiraterone (an androgen synthetic inhibitor) [11], and cabazitaxel (a second-generation taxane) [12] following docetaxel resistance, and their use in CRPC is now becoming widespread. In Japan, cabazitaxel is approved for use in patients following docetaxel resistance [9].

Several studies have reported on the efficacy of cabazitaxel in cancer therapy, with improvements in survival in pretreated men with CRPC in clinical trials [12,13,14] and correspondingly good survival outcomes in real-world practice [15,16,17,18]. One of the most common Grade ≥ 3 adverse events (AEs) associated with the use of cabazitaxel is neutropenia or febrile neutropenia [12, 18, 19] and, interestingly, there appears to be a correlation between efficacy and rates of Grade ≥ 3 neutropenia [20, 21]. Other prognostic factors that have been reported to be associated with clinical outcomes in CRPC patients treated with cabazitaxel include site of metastasis [18, 22,23,24], Eastern Cooperative Oncology Group performance status (ECOG PS) [24, 25], number of cabazitaxel treatment cycles [23], prior treatment history [18, 26] and several laboratory measures [25, 27, 28].

We have previously reported the safety and effectiveness of cabazitaxel in a post marketing surveillance (PMS) of 660 patients with CRPC [15,16,17]. An analysis of outcomes according to cabazitaxel dose suggested that a higher dose may extend overall survival (OS) and time-to-treatment failure (TTF), but it can also induce more events of neutropenia and febrile neutropenia [17]. It is unclear whether the dose of cabazitaxel or the development of neutropenia is the key to predicting survival outcomes. We conducted a network analysis of the data from the PMS to comprehensively evaluate the relationship between clinical factors and patient outcomes using machine learning technology. The objective of this exploratory signal-finding study was to identify correlations between clinical outcomes (response variables) and potential explanatory variables using a network clustering analysis.

Methods

Design of the PMS

The PMS was a prospective, multicenter, observational study, which registered all patients with metastatic, docetaxel-refractory CRPC treated with cabazitaxel following its launch in Japan in September 2014 [15]. The PMS was conducted in compliance with the Ministerial Ordinance on Good Post-marketing Study Practice for Drugs in Japan, was in line with Japanese law, and did not require patient consent for participation in accordance with local regulations and because data were collected anonymously.

Full details of the PMS have been reported [15]. In brief, a total of 660 patients were enrolled across 316 centers by June 2016. In general, cabazitaxel (25 mg/m2) was infused over 1 h every 3 weeks in combination with daily oral prednisolone, in accordance with the approved package insert [29]. Prophylactic granulocyte colony-stimulating factor was recommended for patients susceptible to febrile neutropenia.

Efficacy was assessed in terms of OS, evaluated from the date of first cabazitaxel administration to date of death from any cause; TTF, defined as the duration of cabazitaxel treatment; and PSA response rate, defined as a decrease of ≥ 30% from baseline where the PSA at baseline was ≥ 5 ng/mL.

Adverse drug reactions were evaluated according to Common Terminology Criteria for Adverse Events version 4.0. Effectiveness endpoints (OS, TTF, and PSA response) were assessed for up to 1 year.

Machine learning analysis

Two types of analyses were conducted to identify correlations between patient and disease factors and clinical outcomes. For the first analysis, we conducted graphical Markov (GM) model-based simulations [30, 31] based on the partial correlation coefficient between response and explanatory variables. Response variables included OS, TTF, and PSA response. The 91 explanatory variables were derived from patient demographic and clinical characteristics at baseline, medical and treatment histories, and the AEs collected during the PMS. The choice of factors was based on the bootstrap method and P-value of each variable.

Response and explanatory variables are shown in Additional File 1. Simulations were performed in 1000 iterations, and the variables were considered as correlated when the P-value for the partial correlation coefficient was smaller than the defined alpha level (0.05 or 0.01) in more than 500 iterations. A graphical mixed model using the path consistency algorithm [32] was implemented to avoid interruption of analysis owing to the inability to calculate partial correlation coefficients between the variables. The statistical analysis and machine learning analysis were performed using ‘R’ (version 3.5.1) [33] to analyze and visualize the associations.

For the second analysis, we used network clustering in ‘R’ package (‘R’ version 3.5.1) to define the factors that were associated with the response variable(s) in more than four out of seven clustering models. The clustering models were edge betweenness, eigenvectors of matrices, community structure, fast unfolding of communities, near linear time algorithm, random walks, and maps of random walks [34]. The causality of association between variables was judged to be positive if the frequency of association was positive in 80% of the testing or if explanatory variables were repeatedly clustered into the same group as response variables.

OS analysis by patient subgroup

Factors that were shown by the machine learning analysis to be associated with OS were confirmed according to the clinical outcomes observed in the PMS. Kaplan–Meier methodology was used to visualize the data; calculations were conducted using SAS software version 9.2 or 9.4 (SAS Institute Inc., Cary, NC, USA).

Results

Patients

Patient characteristics and cabazitaxel dosing conditions are shown in Table 1. The median patient age was 70.0 years, 97.9% had previously received docetaxel, 86.5% had received prior androgen receptor inhibitors, and 29.9% had received palliative radiation therapy. A total of 516 patients (78.2%) had a Gleason score of 8–10, and a range of metastatic sites were reported, including bone (88.0%), liver (13.3%), and lung (10.6%). The median number of cycles of cabazitaxel was 4.0 (min–max 1–18), and the median relative dose intensity (RDI) was 67.2% (min–max 17.8–101.0).

Table 1 Patient characteristics and cabazitaxel dosing conditions
Full size table

Treatment effectiveness and safety outcomes are reported in Table 2. Median OS was 319 days (95% confidence interval: 293–361) and median TTF was 116 days (95% confidence interval: 108–135). Neutropenia-associated events occurred in 382 (57.9%) patients, and 325 (49.2%) patients experienced Grade ≥ 3 events. Febrile neutropenia occurred in 119 (18.0%) patients.

Table 2 Treatment effectiveness and safety outcomes
Full size table

GM model-based simulations

Table 3 shows the results of graphical modeling. When the threshold for correlation was set at 0.05 (Additional File 2a and Additional File 3), factors found to be associated with OS included presence of liver metastases, PS, TTF, and neutropenia. At a threshold value of 0.01 (Additional File 2b and Additional File 4), liver metastases, PS, and TTF retained their association but neutropenia did not retain its association. Factors associated with TTF were OS and RDI (threshold 0.05), and OS (threshold 0.01).

Table 3 Results of graphical Markov modeling
Full size table

Network clustering in ‘R’ package using GM model-based simulations

The results of the clustering analysis are shown in Table 4. At a threshold of 0.05 (Additional File 5), the number of treatment cycles, discontinuation due to disease progression, and TTF correlated with OS. In addition, at a threshold of 0.01 (Additional File 6), liver and lung metastases, PS, discontinuation due to AEs and other reasons, and febrile neutropenia were found to be correlated with OS.

Table 4 Results of clustering analysis
Full size table

OS analysis by patient subgroup

The results of the Kaplan–Meier analysis of OS for patient subgroups (Fig. 1) demonstrated that presence of visceral metastases and poor PS at baseline were associated with worse OS. Conversely, development of neutropenia or febrile neutropenia, and a higher number of cabazitaxel cycles were associated with better OS.

Fig. 1
figure 1

Kaplan–Meier analysis of OS for patient subgroups. a: liver metastasis; b: lung metastasis; c: ECOG PS; d: neutropenia; e: febrile neutropenia; f: number of cycles of cabazitaxel administration. CI confidence interval; ECOG PS Eastern Cooperative Oncology Group performance status; HR hazard ratio; OS overall survival

Full size image

Discussion

The definitive identification of prognostic factors for CRPC is critical for improving therapeutic decision-making. It is expected that the use of machine learning algorithms can overcome the limitations of conventional statistical assays, which have low levels of flexibility and performance restrictions due to the number of variables that can be evaluated [3]. Such algorithms can provide a much-needed new and robust information source to guide physicians in evaluating clinical risks and outcomes. This exploratory, signal-finding, machine learning analysis was intended to comprehensively identify factors associated with cabazitaxel in CRPC, using data from a PMS.

Previous analyses have suggested a positive correlation between cabazitaxel efficacy in terms of improved OS and/or progression-free survival and rates of cabazitaxel-induced Grade ≥ 3 neutropenia [20, 21]. These previous studies did not assess the relationship between cabazitaxel dosing and OS benefit. Data from our PMS suggested that a higher dose of cabazitaxel could extend OS but was also associated with an increased incidence of neutropenia [17]. The novelty of our study is that we further analyzed the PMS data to clarify whether the dose of cabazitaxel or the development of neutropenia was the key factor in predicting OS. Although neutropenia was found to be associated with OS, a relationship between the RDI of cabazitaxel and OS was not detected in this analysis, despite RDI in the network analysis being located near febrile neutropenia and neutropenia via TTF. The phase 3 PROSELICA study found that cabazitaxel 20 mg/m2 (reduced dose) was noninferior to cabazitaxel 25 mg/m2 (approved dose) in post-docetaxel patients with metastatic CRPC [14]. In that study, the reduced dose of cabazitaxel maintained ≥ 50% of the OS benefit of the approved dose versus mitoxantrone (data from the phase 3 TROPIC study), thus, the noninferiority endpoint was met. While fewer AEs were observed with the reduced dose, secondary endpoints, including progression-free survival and PSA, favored the approved dose.

In this analysis, factors associated with worse outcomes for men with metastatic CRPC included poor PS and presence of liver/lung metastases. This is in accordance with prior studies that have also identified PS [24, 25] and visceral metastases as indicators of poor prognosis [18, 22,23,24]. We consider that this concordance supports the results of this exploratory machine learning analysis.

Using the GM and clustering methodologies, respectively, neutropenia and febrile neutropenia were identified as factors correlated with clinical outcomes in our analyses. This confirms the significance of this AE in the treatment of CRPC with cabazitaxel, as previously reported [20, 21]. Development of neutropenia during various cancer treatments has also been linked with improved clinical outcomes for patients with other tumor types, suggesting that neutropenia could potentially be a predictive factor for cancer treatment efficacy [35,36,37].

Notably, at the thresholds set in this analysis, cabazitaxel dose and use of granulocyte colony-stimulating factor were not associated with OS. RDI was also not associated with OS. There was a weak and negative correlation between RDI and TTF, although it remains to be determined whether this could have indirectly influenced OS. Overall, these data suggest that neutropenia, rather than the dose-related parameters of cabazitaxel, is associated with OS.

The present study data also indicated correlations with OS for several factors related to the treatment period of cabazitaxel (including TTF, number of cycles, and discontinuation of treatment). Data from a recent retrospective analysis also suggested a link between the number of cabazitaxel treatment cycles and survival. Patients receiving ≥ 4 cycles had significantly longer OS than those who received < 4 cycles (P < 0.001) [24].

There is currently a great deal of interest in machine learning to predict clinical outcomes in oncology [38,39,40]. It is thought that the widespread use of this technique could revolutionize future oncologic management and assist in the implementation of precision medicine [41]. Although some technical refinements are still necessary, the evidential value of the data from this analysis is strengthened and supported by identifying several survival-associated factors detected in prior analyses. Other advantages of this machine learning methodology are the lack of limitation on included factors and inclusion of patients who undergo dose increases. As such, we consider that this methodology may be implemented to analyze a range of real-world data, including registry studies for oncology drugs, to provide physicians with critical information to assist with patient management.

The current analysis has several limitations, one of which is that the observation period of the PMS was limited to 1 year. In this study, there are two types of censored populations. One corresponds to the censored population within 365 days and the other to the population observed throughout the 365 days, after which observation was stopped. Both populations were labeled with the same variable, and the techniques used for handling missing outcomes and censored cases may have introduced bias into the results. In addition, no association between PSA response and outcome was observed in this analysis, possibly due to a partial lack of data and the difficulty of categorizing baseline PSA levels as parameters. A relationship between the dose of cabazitaxel and OS was not detected in the present study. This does not negate our previous findings that the initial cabazitaxel dose exerted an effect on the clinical outcomes; this may have been because we used a trimmed dataset for the calculation (the dataset of the previous analysis did not include patients who received an escalated dose of cabazitaxel from the initial dose), and because physicians tend to use the most appropriate dose for each patient. Additionally, the cabazitaxel dose is directly associated with the complications of infection or bone suppression [17], possibly because longer treatment increases the chance of a treatment-emergent AE, which would be a time-dependent factor. There may also have been limitations related to the explanatory parameters, as these sometimes represent a combination of clinically significant factors. For example, TTF may be due to an AE or progressive disease. Finally, it must be noted that this was an exploratory analysis, intended to identify signals, rather than a model validation study to prove prognostic relationships. We aimed to prove the value of machine learning technology in the assembly of a model that could be used to evaluate correlative associations; our study was not designed to prove causative pathophysiologic associations between patient or disease variables and the subsequent clinical outcomes. The correlation analysis was completed at model creation without further verification, and no definitive predictive value can be inferred. Our machine learning model did indicate several avenues of interest to be explored in terms of the relationships between variables and outcomes, and many of the results were consistent with those in the published literature. However, further studies will be required to validate such models with additional cohort data, and confirm or disprove hypotheses relating to prognostic ability and the attribution of causation.

Conclusions

This analysis suggests that neutropenia may be correlated with treatment efficacy in terms of survival. Poor PS and distant metastases to the liver and lungs were determined to be associated with worse outcomes. In contrast, factors related to treatment duration were shown to positively correlate with improved OS. The identification of factors that have been previously reported to be associated with survival supports the results of our machine learning analysis and strengthens the value of this technique as a potentially powerful tool in the assessment and analysis of clinical risks and outcomes.