Abstract
Background:
Accurate risk stratification is critical to guide management decisions in localized prostate cancer (PCa). Previously, we had developed and validated a multimodal artificial intelligence (MMAI) model generated from digital histopathology and clinical features. Here, we externally validate this model on men with high-risk or locally advanced PCa treated and followed as part of a phase 3 randomized control trial.
Objective:
To externally validate the MMAI model on men with high-risk or locally advanced PCa treated and followed as part of a phase 3 randomized control trial.
Design, setting, and participants:
Our validation cohort included 318 localized high-risk PCa patients from NRG/RTOG 9902 with available histopathology (337 [85%] of the 397 patients enrolled into the trial had available slides, of which 19 [5.6%] failed due to poor image quality).
Outcome measurements and statistical analysis:
Two previously locked prognostic MMAI models were validated for their intended endpoint: distant metastasis (DM) and PCa-specific mortality (PCSM). Individual clinical factors and the number of National Comprehensive Cancer Network (NCCN) high-risk features served as comparators. Subdistribution hazard ratio (sHR) was reported per standard deviation increase of the score with corresponding 95% confidence interval (CI) using Fine-Gray or Cox proportional hazards models.
Results and limitations:
The DM and PCSM MMAI algorithms were significantly and independently associated with the risk of DM (sHR [95% CI] = 2.33 [1.60–3.38], p < 0.001) and PCSM, respectively (sHR [95% CI] = 3.54 [2.38–5.28], p < 0.001) when compared against other prognostic clinical factors and NCCN high-risk features. The lower 75% of patients by DM MMAI had estimated 5- and 10-yr DM rates of 4% and 7%, and the highest quartile had average 5- and 10-yr DM rates of 19% and 32%, respectively (p < 0.001). Similar results were observed for the PCSM MMAI algorithm.
Conclusions:
We externally validated the prognostic ability of MMAI models previously developed among men with localized high-risk disease. MMAI prognostic models further risk stratify beyond the clinical and pathological variables for DM and PCSM in a population of men already at a high risk for disease progression. This study provides evidence for consistent validation of our deep learning MMAI models to improve prognostication and enable more informed decision-making for patient care.
Patient summary:
This paper presents a novel approach using images from pathology slides along with clinical variables to validate artificial intelligence (computer-generated) prognostic models. When implemented, clinicians can offer a more personalized and tailored prognostic discussion for men with localized prostate cancer.
Keywords: Artificial intelligence, Biomarker, Digital histopathology, Prostate cancer
1. Introduction
Prostate cancer (PCa) is the most common malignancy in men, with a lifetime risk of diagnosis approaching nearly one in nine [1]. PCa is heterogeneous in nature, including both latent and clinically aggressive disease progression pathways [2]. With routine screening, PCa is primarily discovered at earlier stages of disease, leading to concerns for overtreatment among men with clinically indolent disease [2]. In contrast, there is also a significant risk of undertreatment among men with high-risk tumors, where definitive treatment can offset the risk of symptomatic progression of PCa and death [3]. Accurate risk stratification is critical to guide management decisions. To date, prognostic clinicopathological features such as histological Gleason grade, prostate-specific antigen (PSA), and T stage demonstrate poor to modest prognostic ability in identifying disease that is likely to progress [4–7]. In this context, there is often a nuanced and iterative decision-making process between the physician and the patient/family about if and when to initiate definitive treatment for PCa. There is immense opportunity to further refine this process through well-performing prognostic and predictive biomarkers for PCa [7].
Tumor morphology representing the loss of glandular structure and organization of the tissue is associated with the aggressiveness of the disease trajectory for PCa [8]. Gleason scores do not always adequately capture other histological markers of aggressiveness of the disease such as tumor heterogeneity, microenvironment, and stromal architecture of the tumor [9–12]. Artificial intelligence (AI) models based on digital histopathology can learn both high- and low-risk features from large amounts of data for men with available pathology and known outcomes [13,14]. In particular, when digital AI models are developed and validated from pathology associated with clinical trial data, these yield a unique opportunity in modeling within longer event horizons and prediction of clinically meaningful endpoints [15].
We previously described multimodal artificial intelligence (MMAI) models using digital histopathology and clinical data that were developed and validated from five phase 3 PCa trials (NRG/RTOG 9202, 9408, 9413, 9910, and 0126), which outperformed standard risk models (National Comprehensive Cancer Network [NCCN]) in the prediction of distant metastasis (DM) and PCa-specific mortality (PCSM) [13]. In this study, we externally validate our MMAI algorithms for the prediction of DM and PCSM in NRG/RTOG 9902, a phase 3 [16] randomized trial of men with localized or locally advanced high-risk PCa. We also explored our MMAI algorithms’ ability to predict secondary endpoints such as biochemical failure (BF) and overall survival (OS).
2. Patients and methods
2.1. Patients
NRG/RTOG-9902 [16] enrolled 397 high-risk, localized PCa patients who were randomized to receive long-term androgen suppression (AS) with radiotherapy (RT) alone (AS + RT) or with adjuvant combination chemotherapy (CT; AS + RT + CT) between January 2000 and October 2004 [16]. The CT regimen was four 21-d cycles with paclitaxel, estramustine, and oral etoposide delivered beginning 28 d following 70.2 Gy RT. The AS regimen was luteinizing hormone-releasing hormone for 24 mo beginning 2 mo prior to RT plus oral antiandrogen for 4 mo before and during RT. Men were enrolled if they either had PSA between 20 and 100 ng/ml and Gleason score ≥7 or had clinical stage ≥T2 and Gleason score ≥8. Importantly, 10-yr results showed no statistically significant difference in OS, BF, local progression, DM, or disease-free survival between the two treatment arms (p > 0.05 for all endpoints listed). As such, in the present study, we pooled all men as one cohort regardless of the treatment arm.
The current study of the external validation of MMAI models had approval from NRG Oncology, a National Clinical Trials Network (NCTN) group funded by the National Cancer Institute (NCI).
2.2. Sample process and scanning
All available pretreatment biopsy slides from NRG/RTOG-9902 were digitized by NRG Oncology using a Leica Biosystems Aperio AT2 digital pathology scanner at 20× resolution. The histopathology images were reviewed for quality and clarity by the NRG Biobank operator and by the AI data intake team. A previously built artifact classifier was used to filter out low-quality images [13].
2.3. Description of the MMAI model
As described previously, six MMAI algorithms were developed and validated using five phase 3 NRG trials (RTOG 9202, 9408, 9413, 9910, and 0126) utilizing both digital histopathology slides and clinical data from each patient [13]. The MMAI architecture was then updated with a few key improvements, including (1) different development and validation with additional patients added and new stratified random split implemented, (2) limiting to pretreatment tissue patches used to generate the 128-dimension feature vectors, (3) improved multimodal learning with a multiple instance learning–based neural network, (4) use of an attention mechanism with the time to event of the desired clinical endpoints as label, and (5) transurethral resection of the prostate samples excluded during initial validation in 20% of patients from the five trials (see additional model development details in the Supplementary material and Supplementary Fig. 1). From this MMAI architecture (Artera Prostate Prognostic Model version 1.1), there were two locked MMAI algorithm risk scores optimized for the desired clinical endpoints—DM and PCSM. The PCSM MMAI was optimized during training to predict the risk of biologically related PCa death defined as death with DM (DDM). Both algorithms’ scores range from 0 to 1; a higher score indicates a greater risk of experiencing the event. The algorithm score generation process was blinded from the clinical endpoint information.
2.4. Endpoints
The primary endpoints for this study were (1) time to DM, defined as the years from randomization to the date of DM, and (2) time to PCSM, defined as the years from randomization to the date of death due to treated cancer or complications of protocol treatment in the study.
The secondary endpoints were time to BF, defined as the years from randomization to the date of BF (the first of either PSA failure or the initiation of salvage hormone therapy); time to DDM, defined as the years from randomization to the date of death with a record of metastasis; and OS, defined as the years from randomization to the date of death.
Patients who were lost to follow-up before experiencing an event of interest were censored at their last follow-up. For DM, PCSM, DDM, and BF, death before experiencing the event of interest was considered competing events.
2.5. Statistical analysis
The digital pathological evaluable population (DPEP) was defined as individuals who were randomized to RTOG-9902, with relevant study endpoint data, quality histopathology data, and baseline clinical variables including age, clinical T stage, Gleason information (primary, secondary, and total score), and PSA available to generate an MMAI algorithm score for analysis. For MMAI score generation, missing baseline clinical variables (age; PSA; Gleason primary, secondary, and total scores; and T stage) were considered numeric and imputed using sklearn KNNImputer with the default parameters (average of five nearest neighbors) [17]. Complete cases on baseline clinical variables were used for the analysis, including descriptive summary and univariable/multivariable regressions.
The baseline demographic and clinical characteristics were summarized descriptively for the DPEP and intention-to-treat (ITT) population, and compared between the DPEP and the subgroup of patients from the ITT population but without quality histopathology data. Descriptive summaries were provided using count and portion (%) for categorical variables, and median and interquartile range (IQR) for continuous variables. To assess whether the DPEP was representable of the original ITT population, p values were calculated using Wilcoxon rank sum test for continuous variables, and Pearson’s chi-square test or Fisher’s exact test for categorical variables. Median follow-up time was calculated using the reverse Kaplan-Meier method.
The prognostic performance of the MMAI algorithms was assessed using univariable and multivariable analyses. Individual clinical risk factors that are commonly known prognosticators (age, log 2 baseline PSA, total Gleason score in two categories 7 and 8–10, clinical T stage in two categories T1–2 and T3–4, and number of NCCN high-risk features in three categories 1, 2, and 3–4) were adjusted in the multivariable analysis to assess whether MMAI algorithms would remain prognostic consistently while acknowledging the potential risk of overadjustment. Baseline PSA was transformed by taking log base 2 for ease of interpretation. Fine and Gray regression [18] was used to estimate subdistribution hazard ratio (sHR) and 95% confidence interval (CI) for DM, PCSM, and BF endpoints, while Cox proportional hazards regression was used to estimate the hazard ratio (HR) and 95% CI for OS. The MMAI algorithm scores were explored by quartile splits; specifically, quartiles with similar prognoses were grouped together and summarized using cumulative incidence curves [19], with 5- and 10-yr estimated DM and PCSM rates and corresponding two-sided 95% CIs provided. The tests for the MMAI-treatment interaction were also performed as an exploratory analysis.
All statistical analyses were performed using R, version 4.1.2 (R Foundation for Statistical Computing, Vienna, Austria). All statistical tests were two sided and used a 0.05 significance level. The prognostic model validation findings were reported using the TRIPOD reporting criteria [20].
3. Results
3.1. Participants
The MMAI algorithm scores were generated for 318 of the 397 original clinical trial patients enrolled in NRG/RTOG-9902 (337 [85%] of the full trial cohort had available slides, of which 19 (5.6%) failed to be included due to poor image quality). Figure 1 demonstrates the flow of patients from the NRG/RTOG 9902 clinical trial to the DPEP included for model validation. The baseline characteristics of the study DPEP are shown in Table 1. The evaluable population included men with median baseline PSA 23.0 ng/ml, 32% of whom had cT3–4 disease, 67% had grade group 4 or 5 disease, and 54% had more than one NCCN high-risk features. There are no statistically significant differences in baseline characteristics between patients for whom MMAI algorithm scores could be obtained compared with the 62 patients who were excluded from this validation study. Similarly, there were no statistically significant differences in baseline characteristic between the two treatment arms within the DPEP. With a median follow-up of 10.4 yr, 42 men had experienced DM and 29 PCSM. The median (IQR) score of the algorithm optimizing for DM (DM MMAI) was 0.54 (0.44–0.62), and it was 0.53 (0.47–0.60) for the algorithm optimized for PCSM (PCSM MMAI). There were no statistically significant differences for either score between the two treatment arms of NRG/RTOG 9902 (Supplementary Table 1).
Fig. 1 –
Consort flow diagram of patients toward study inclusion from the parent clinical trial NRG/RTOG 9902.
AS = androgen suppression; CT = chemotherapy; DPEP = digital pathological evaluable population; H&E = hematoxylin and eosin; MMAI = multimodal artificial intelligence; RT = radiation therapy.
Table 1 –
Population characteristics by the presence of MMAI classifier status
| Characteristic | RTOG 9902 clinically eligible with follow-up (N = 380) | RTOG 9902 DPEP (N = 318) | |||||
|---|---|---|---|---|---|---|---|
| Overall (N = 380)a | DPEP (N = 318)a | Image not available (N = 62)a | p value | RT + ADT (N = 161)a | RT + ADT + CT (N = 157)a | p value | |
| Age, median (IQR) | 66 (60, 71) | 66 (60, 71) | 66 (60, 71) | 0.8b | 66 (59, 72) | 67 (62, 71) | 0.15b |
| Race, n (%) | 0.4c | 0.5c | |||||
| Black | 102 (27) | 83 (26) | 19 (31) | 46 (29) | 37 (24) | ||
| White | 265 (70) | 222 (70) | 43 (69) | 107 (66) | 115 (73) | ||
| Other | 12 (3.2) | 12 (3.8) | 0 (0) | 7 (4.3) | 5 (3.2) | ||
| Unknown | 1 (0.3) | 1 (0.3) | 0 (0) | 1 (0.6) | 0 (0) | ||
| Zubrod, n (%) | 0.12d | 0.14d | |||||
| 0 | 340 (89) | 288 (91) | 52 (84) | 142 (88) | 146 (93) | ||
| 1 | 40 (11) | 30 (9.4) | 10 (16) | 19 (12) | 11 (7.0) | ||
| Baseline PSA (ng/ml) | 0.5b | 0.7b | |||||
| Median (IQR) | 23 (9, 40) | 23 (9, 41) | 22 (8, 39) | 23 (10, 40) | 23 (9, 45) | ||
| 0.5d | 0.6d | ||||||
| <20, n (%) | 164 (43) | 135 (42) | 29 (47) | 66 (41) | 69 (44) | ||
| 20+, n (%) | 216 (57) | 183 (58) | 33 (53) | 95 (59) | 88 (56) | ||
| Gleason, n (%) | 0.4d | 0.7d | |||||
| 7 | 125 (33) | 105 (33) | 20 (32) | 54 (34) | 51 (32) | ||
| 8 | 135 (36) | 109 (34) | 26 (42) | 58 (36) | 51 (32) | ||
| 9–10 | 120 (32) | 104 (33) | 16 (26) | 49 (30) | 55 (35) | ||
| Clinical T stage, n (%) | 0.9d | 0.4d | |||||
| T1-T2a | 144 (38) | 121 (38) | 23 (37) | 67 (42) | 54 (35) | ||
| T2b-c | 111 (29) | 94 (30) | 17 (27) | 45 (28) | 49 (31) | ||
| T3-T4 | 124 (33) | 102 (32) | 22 (35) | 49 (30) | 53 (34) | ||
| Missing, n | 1 | 1 | 0 | 0 | 1 | ||
| Grade group, n (%) | 0.6d | 0.8d | |||||
| 2 | 62 (16) | 51 (16) | 11 (18) | 25 (16) | 26 (17) | ||
| 3 | 63 (17) | 54 (17) | 9 (15) | 29 (18) | 25 (16) | ||
| 4 | 135 (36) | 109 (34) | 26 (42) | 58 (36) | 51 (32) | ||
| 5 | 120 (32) | 104 (33) | 16 (26) | 49 (30) | 55 (35) | ||
| NCCN risk group, n (%) | 0.2c | >0.9c | |||||
| Intermediate | 1 (0.3) | 0 (0) | 1 (1.6) | 0 (0) | 0 (0) | ||
| High/very high | 379 (100) | 318 (100) | 61 (98) | 161 (100) | 157 (100) | ||
| No. of NCCN high-risk features, n (%) | 0.4c | 0.8c | |||||
| 0 | 1 (0.3) | 0 (0) | 1 (1.6) | ||||
| 1 | 175 (46) | 145 (46) | 30 (48) | 74 (46) | 71 (46) | ||
| 2 | 143 (38) | 122 (39) | 21 (34) | 61 (38) | 61 (39) | ||
| 3 | 51 (14) | 42 (13) | 9 (15) | 23 (14) | 19 (12) | ||
| 4 | 7 (1.9) | 6 (1.9) | 1 (1.6) | 2 (1.3) | 4 (2.6) | ||
| Missing, n | 3 | 3 | 0 | 1 | 2 | ||
| Distant metastasis, n (%) | 50 (13) | 42 (13) | 8 (13) | >0.9d | 23 (14) | 19 (12) | 0.6d |
| Prostate cancer–specific mortality, n (%) | 36 (9.5) | 30 (9.4) | 6 (9.7) | >0.9d | 15 (9.3) | 15 (9.6) | >0.9d |
| Follow-up for censored Pts (yr), median (IQR) | 10 (9, 12) | 10 (9, 12) | 10 (9, 11) | 0.5b | 10 (9, 12) | 10 (9, 12) | 0.3b |
ADT = androgen deprivation therapy; CT = chemotherapy; DPEP = digital pathological evaluable population; IQR = interquartile range; MMAI = multimodal artificial intelligence; NCCN = National Comprehensive Cancer Network; Pts = patients; RT = radiotherapy.
Some percentages may not add up to a hundred percent due to rounding.
Wilcoxon rank sum test.
Fisher’s exact test.
Pearson’s chi-square test.
3.2. Model performance
Compared with clinical and pathological factors, the MMAI algorithms were significantly prognostic across outcome measures. In the univariable analysis, the DM MMAI algorithm continuous score was statistically associated with the DM endpoint (sHR 2.33, 95% CI 1.60–3.38, p < 0.001), and the PCSM MMAI algorithm was associated with the PCSM endpoint (HR 3.54, 95% CI 2.38–5.28, p < 0.001; Table 2). When evaluating the secondary endpoints, DM MMAI was statistically significantly associated with the risks of BF, DDM, PCSM, and OS. Similarly, PCSM MMAI was statistically significantly associated with the risk of DM, DDM, and death of any cause (Supplementary Fig. 2).
Table 2 –
Univariable analysis of association between MMAI algorithms and DM and PCSM endpoints
| Variables | Levels | DM MMAI for DM endpoint | PCSM MMAI for PCSM endpoint | ||
|---|---|---|---|---|---|
| sHR (95% CI) | p value | sHR (95% CI) | p value | ||
| Age | 0.97 (0.93–1.00) | 0.05 * | 0.97 (0.92–1.01) | 0.16 | |
| Log2 baseline PSA | 0.89 (0.70–1.14) | 0.37 | 0.60 (0.45–0.81) | <0.001 * | |
| Gleason score | 8–10 vs 3 + 4 | 1.14 (0.30–4.32) | 0.85 | 0.92 (0.18–4.58) | 0.91 |
| 4 + 3 vs 3 + 4 | 1.95 (0.67–5.67) | 0.22 | 1.93 (0.57–6.58) | 0.29 | |
| T stage | T3-T4 vs T1-T2 | 1.48 (0.80–2.75) | 0.21 | 2.51 (1.22–5.14) | 0.01 * |
| CAPRA risk groups | 6–10 vs 3–5 | 0.76 (0.33–1.75) | 0.52 | 0.52 (0.22–1.23) | 0.14 |
| MMAI (per SD increase) | 2.33 (1.60–3.38) | <0.001 * | 3.54 (2.38–5.28) | <0.001 * | |
CI = confidence interval; DM = distant metastasis; MMAI = multimodal artificial intelligence; PCSM = prostate cancer–specific mortality; PSA = prostate-specific antigen; SD = standard deviation; sHR = subdistribution hazard ratio.
The prognostic effect size for DM for the MMAI model was consistent across subgroups, including treatment arm, age, race, PSA, clinical T stage, Gleason score, and number of high-risk features (Fig. 2A). Similarly, for the PCSM endpoint, the PCSM MMAI was also prognostic in all subgroups (Fig. 2B). In the multivariable analysis (Fig. 3 and Supplementary Table 2), controlling for individual clinical factors including age, log 2 baseline PSA, total Gleason score, clinical T stage, and number of NCCN high-risk factors, both DM MMAI and PCSM MMAI were consistently significant prognosticators.
Fig. 2 –
Prognostic performance of MMAI models for (A) distant metastasis (DM) and (B) prostate cancer–specific mortality (PCSM) within subgroup classifications.
ADT = androgen deprivation therapy; CI = confidence interval; CT = chemotherapy; MMAI = multimodal artificial intelligence; NCCN = National Comprehensive Cancer Network; PSA = prostate-specific antigen; RT = radiation therapy; sHR = subdistribution hazard ratio.
Fig. 3 –
Multivariable analysis of MMAI algorithms on (A) DM and (B) PCSM after adjusting for individual clinical risk factors.
CI = confidence interval; DM = distant metastasis; MMAI = multimodal artificial intelligence; MVA = multivariable analysis; PCSM = prostate cancer–specific mortality; PSA = prostate-specific antigen; sHR = subdistribution hazard ratio.
DM MMAI was explored by dividing into quartiles and based on the 5- and 10-yr estimated DM rates; the first three quartiles (Q1–3) were grouped together, similarly for PCSM MMAI (Supplementary Fig. 3). The lower 75% (Q1–3) of patients had estimated 5- and 10-yr DM rates of 4% (95% CI 1–6%) and 7% (95% CI 4–10%), respectively, and the highest quartile (Q4) had estimated 5- and 10-yr DM rates of 19% (95% CI 10–28%) and 32% (95% CI 21–43%), respectively, with an sHR of 5.1 (95% CI 2.7–9.3, p < 0.001; Fig. 4A). Similar results were observed for the PCSM MMAI (sHR 5.9, 95% CI 2.8–12.2, p < 0.001; Fig. 4B).
Fig. 4 –
Cumulative incidence curves for (A) estimated distant metastasis (DM) risk by quartile 4 versus quartile 1–3 multimodal artificial intelligence optimized for DM (DM MMAI) and (B) estimated prostate cancer–specific mortality risk (PCSM) by quartile 4 versus quartile 1–3 multimodal artificial intelligence optimized for PCSM (PCSM MMAI).
CI = confidence interval; Est. = estimated; MMAI = multimodal artificial intelligence; sHR = subdistribution hazard ratio.
No statistically significant interaction was found either between DM MMAI quartile groups (Q4 vs Q1–3) and CT treatment effect (interaction p = 0.08), or between PCSM MMAI quartile groups and CT treatment effect (interaction p = 0.58; Supplementary Fig. 4). Among the top 25% of patients ranked by DM MMAI, the estimated 5-yr absolute benefit from additional use of CT was 14%, and the 10-yr benefit was 18% (Supplementary Fig. 4A).
4. Discussion
The MMAI prognostic models were originally trained and validated from five phase 3 randomized trials, and have now been successfully validated independently in the NRG/RTOG 9902 randomized trial. The present study demonstrated that the differences between MMAI algorithm lower- versus higher-risk groups for DM and PCSM were large and statistically significant even within this NCCN high- and very-high-risk population. In multivariable analyses, the MMAI score was independently prognostic, even after controlling for variables known to be associated with prognostic risk (patient age, baseline PSA, total Gleason score, and T stage). Association of the MMAI algorithms with DM and PCSM within subgroups suggested added discrimination and prognostic ability of MMAI throughout the continuum of high- and very-high-risk disease.
To date, several novel biomarkers have been developed for localized PCa, which have been shown to outperform clinical pathological factors alone [7,20–22]. Despite this, there are very few biomarkers that are integrated into routine clinical practice [22]. A challenge for the generation of novel biomarkers in localized PCa is multifactorial. This includes the need for high-quality training data with rigorous and long-term annotation to enable the collection of clinically meaningful endpoints, such as DM and PCSM. Furthermore, large sample sizes are needed given the generally low event rate in PCa compared with other cancer types, such as pancreatic cancer. Diverse populations are also important to ensure generalizability of biomarkers. Similar requirements exist for validation. The NRG Oncology Biobank, combined with the advantages of AI digital histopathology radiomics, has enabled the generation of multiple MMAI prognostic and predictive models that leverage the many strengths of large prospective trials that have enrolled patients across hundreds of centers, include diverse patient populations, contain large sample sizes, and utilize a biomarker that does not consume tissue. Further, the MMAI algorithms allow for rapid turnaround time and require no additional tissue processing, which could be very advantageous in its dissemination as a prognostic tool for clinicians treating men with PCa. The use of an MMAI algorithm that includes self-supervised learning from both digital histopathology and clinical characteristics offers a novel approach that is scalable in various multiple clinical environments, including clinical settings that are resource constrained. Further development of predictive classifiers can be used to personalize treatment decisions and be integrated in clinical care environments to offer guidance on nuanced treatment decisions. This work extends the literature supporting the potential utility of digital histopathology AI models in clinical practice, as suggested by the NCCN [23].
There were numerous strengths of this approach, including the use of the prospective NRG/RTOG 9902 clinical trial with systematically adjudicated outcome events and long-term follow-up. Further, as the trials included NCCN high- and very-high-risk men, events of interest are clinically meaningful (DM and PCSM). Limitations of this study include that the included Gleason information is based on original classification reported and has not been reclassified by current grading definitions. There are neither any genomic biomarkers nor clinical nomograms available as comparators in this study due to data limitation. Future work assessing MMAI performance in more homogeneous risk populations and subgroups of interest is needed. Similarly, due to the era in which the trial was conducted, men were not diagnosed and staged with the use of prostate-specific membrane antigen positron emission tomography and multiparametric magnetic resonance imaging, there have been multiple changes in treatment and imaging in routine clinical care. While these may impact absolute event rates, these are unlikely to impact the prognostic superiority of the MMAI models over routine clinicopathological variables. Further, prognostic signals based on the MMAI models are developed agnostic to treatment selection; this study is underpowered to investigate the MMAI models for response to CT.
5. Conclusions
Previously locked and validated MMAI prognostic models outperform clinical and pathological variables for the prediction of DM and PCSM in a population of men at a high risk for disease progression. This study provides important evidence for consistent validation of our deep learning MMAI models to improve prognostication and enable more informed and personalized decision-making for patient care. This MMAI model is now incorporated within NCCN guidelines for prognostic risk stratification [23]. Further study is warranted to understand the benefits of integration of the MMAI model within routine clinical care environments.
Supplementary Material
Multimodal artificial intelligence models using digital histopathology slides outperform clinical and pathological variables for prognostic prediction of distant metastasis and prostate cancer–specific mortality, and can be incorporated in clinical practice for personalized risk stratification.
Funding/Support and role of the sponsor:
This project was supported by grants U10CA180868, U10CA180822, UG1CA189867, and U24CA196067 from the National Cancer Institute and Pfizer and Bristol Myers Squibb.
Financial disclosures:
Ashley E. Ross certifies that all conflicts of interest, including specific financial interests and relationships and affiliations relevant to the subject matter or materials discussed in the manuscript (eg, employment/affiliation, grants or funding, consultancies, honoraria, stock ownership or options, expert testimony, royalties, or patents filed, received, or pending), are the following: Kwang Choi, Michael C. Dobelbower, Christopher U. Jones, Scott McGinnis, Jeff M. Michalski, Osama Mohamad, Todd M. Morgan, Stephanie L. Pugh, Seth A. Rosenthal, Ashley E. Ross, Edward M. Schaeffer, and Kenneth L. Zeitzer have nothing to declare. Emmalyn Chen, Huei-Chung Huang, Jingbin Zhang, Andre Esteva, Jessica Keim-Malpass, and Rikiya Yamashita declare that they work at Artera. Robert T. Dess declares, in the past 36 mo, consulting fees from Janssen, unrelated to the current work. Sandy DeVries declares, since the initial planning of the work, funding from the NCI Grant # U24CA196067 to NRG Oncology/UCSF supporting NRG Biospecimen Bank. Felix Y. Feng declares, in the past 36 mo, consulting for Janssen Oncology, Astellas Pharma, Serimmune, Foundation Medicine, Exact Sciences, Bristol-Myers Squibb, and Varian Medical Systems (termed); stock options from Artera (role: medical advisor); and receiving stock options for serving on their advisory boards from BlueStar Genomics and SerImmune. Leonard G. Gomella declares, since the initial planning of the work, being an RTOG/NRG investigator; in the past 36 mo, advisory boards/consulting fees from Merck, Astra Zeneca, Lantheu, and Janssen; and patents planned, issued, or pending Shed Cell Patents owned by Jefferson and SUO Board of Directors. Alan C. Hartford declares, in the past 36 mo, the following: payment for travel to annual ACR meetings from American Society for Radiation Oncology (ASTRO); payment for travel to biannual AMA HOD meetings from New Hampshire Medical Society (NHMS); President (volunteer position) of the Council of Affiliated Regional Radiation Oncology Societies (CARROS), American College of Radiology; Councilor (volunteer position) representing Council American Society for Radiation Oncology (ASTRO) to ACR; and Alternate Delegate (volunteer position) of New Hampshire Medical Society (NHMS) to AMA House of Delegates. Jessica Keim-Malpass declares, in the past 36 mo, stock options related to Artera. Adam Raben declares, in the past 36 mo, being a member of the NRG GU Steering Committee. Howard M. Sandler declares, since the initial planning of the work, support for the role as NRG Oncology GU Committee Chair; in the past 36 mo, role in Clinical Trial Steering Committee and being a member of Janssen; and being a member of ASTRO Board of Directors. A. Oliver Sartor declares the following: in the past 36 mo, grants or contracts from Advanced Accelerator Applications(AAA), Amgen, AstraZeneca, Bayer, Constellation, Endocyte, Invitae, Janssen, Lantheus, Merck, Progenics, and Tenebio; consulting fees from AAA, Amgen, ArtBio, Astellas, AstraZeneca, Bayer, Blue Earth Diagnostics, Inc., Clarity Pharmaceuticals, Clovis, Constellation, Convergent, Dendreon, EMD Serono, Foundation Medicine, Fusion, Genzyme, Hengrui, Isotopen Technologien Meunchen, Merck, Janssen, Morphimmune, Myovant, Myriad, Noria Therapeutics, Inc., NorthStar, Novartis, Noxopharm, Progenics, POINT Biopharma, Pfizer, Sanofi, Tenebio, Telix, Tessa, and Theragnostics; payment or honoraria from AAA, Amgen, ArtBio, Astellas, AstraZeneca, Bayer, Clarity Pharmaceuticals, Convergent, Elsevier Foundation Medicine, Fusion, Genzyme, Hengrui, Isotopen Technologien Meunchen, Merck, Janssen, Morphimmune, NorthStar, Novartis, Propella POINT Biopharma, Pfizer, Sanofi, Tenebio, Telix, Tessa, and Theragnostics; payment for expert testimony from Sanofi; Koochekpour, Sartor AO, inventors: Saposin C and receptors as targets for treatment of benign and malignant disorders; US patent awarded on January 23, 2007 (patent no. 7,166,691); participation in a data safety monitoring board or advisory board for Pfizer, JNJ, AstraZeneca, Memorial Sloan Kettering, Eurtoc, and Merck; and stock or stock options in Clarity Pharmaceuticals, Noria Therapeutics, Inc., Lilly, Clovis, GlaxoSmithKline, Abbvie, Cardinal Health, and United Health Group. Jeffry P. Simko declares, since the initial planning of the work, funding from the grant support paid to own institution from NCI; in the past 36 mo, leadership or fiduciary role for NRG Oncology: Pathology Committee Chair, Research Strategy Committee Member, and Biobank Associate Director; and stock from Protean Biosciences Inc., Alpenglow Biosciences Inc., and Triopsy Medical Inc. Luis Souhami declares, in the past 36 mo, grants or contracts and payment or honoraria from Varian Medical Systems. Daniel E. Spratt declares, in the past 36 mo, grants or contracts from Janssen paid to institution; consulting fees for Hydrogel spacer; consulting for Boston Scientific; and personal fees for advisory boards from Astellas, AstraZeneca, Elekta, Varian, Gamma Tile, Bayer, Janssen, Novartis, Myovant, and Pfizer. Phuoc T. Tran declares, in the past 36 mo, patent licensed to company from Natsar Pharmaceuticals; consulting fees from SAB of RefleXion Medical Inc, Natsar Pharmaceuticals, and Bayer Healthcare & Janssen; Vice Chair of Research NRG Oncology AACR Cancer Research, Journal of Clinical Oncology; University of Maryland SOM Radiation Oncology Vice Chair; NRG GU TS Chair; member of AACR RSM Committee; CS Senior Editor; and member of JCO Editorial Board. Rikiya Yamashita declares, in the past 36 mo, stock options related to Artera. Jingbin Zhang declares, in the past 36 mo, stock options related to Artera.
Footnotes
Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
Data sharing:
All data will be made available as per the NCTN Data Archive rules. The link for the archive is the following: https://nctn-data-archive.nci.nih.gov/.
References
- [1].Siegel RL, Miller KD, Fuchs HE, Jemal A. Cancer statistics, 2022. CA Cancer J Clin 2022;72:7–33. [DOI] [PubMed] [Google Scholar]
- [2].Narayan V, Jiang S, Warlick CA. Early stage cancer in older adults: prostate-avoiding overtreatment and undertreatment. Cancer J 2017;23:238–41. [DOI] [PubMed] [Google Scholar]
- [3].Lichtensztajn DY, Leppert JT, Brooks JD, et al. Undertreatment of high-risk localized prostate cancer in the California Latino population. J Natl Compr Canc Netw 2018;16:1353–60. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [4].Feng FY, Huang H-C, Spratt DE, et al. Validation of a 22-gene genomic classifier in patients with recurrent prostate cancer: an ancillary study of the NRG/RTOG 9601 randomized clinical trial. JAMA Oncol 2021;7:544–52. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [5].Yu EY, Lin DW. Avoiding undertreatment of aggressive prostate cancer by early use of chemotherapy. JAMA Oncol 2017;3:13–4. [DOI] [PubMed] [Google Scholar]
- [6].Muralidhar V, Zhang J, Wang Q, et al. Genomic validation of 3-tiered clinical subclassification of high-risk prostate cancer. Int J Radiat Oncol Biol Phys 2019;105:621–7. [DOI] [PubMed] [Google Scholar]
- [7].Farha MW, Salami SS. Biomarkers for prostate cancer detection and risk stratification. Ther Adv Urol 2022;14:17562872221103988. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [8].Gleason DF, Mellinger GT. Prediction of prognosis for prostatic adenocarcinoma by combined histological grading and clinical staging. J Urol 1974;111:58–64. [DOI] [PubMed] [Google Scholar]
- [9].Leo P, Janowczyk A, Elliott R, et al. Computer extracted gland features from H&E predicts prostate cancer recurrence comparably to a genomic companion diagnostic test: a large multi-site study. NPJ Precis Oncol 2021;5:35. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [10].Zhou M, Li J, Cheng L, et al. Diagnosis of “poorly formed glands” Gleason pattern 4 prostatic adenocarcinoma on needle biopsy: an interobserver reproducibility study among urologic pathologists with recommendations. Am J Surg Pathol 2015;39:1331–9. [DOI] [PubMed] [Google Scholar]
- [11].Hassan O, Matoso A. Clinical significance of subtypes of Gleason pattern 4 prostate cancer. Transl Androl Urol 2018;7:S477–83. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [12].Wasinger G, Oszwald A, Shariat SF, Compérat E. Histological patterns, subtypes and aspects of prostate cancer: different aspects, different outcomes. Curr Opin Urol 2022;32:643–8. [DOI] [PubMed] [Google Scholar]
- [13].Esteva A, Feng J, van der Wal D, et al. Prostate cancer therapy personalization via multi-modal deep learning on randomized phase III clinical trials. NPJ Digit Med 2022;5:71. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [14].Esteva A, Chou K, Yeung S, et al. Deep learning-enabled medical computer vision. NPJ Digit Med 2021;4:5. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [15].Shah P, Kendall F, Khozin S, et al. Artificial intelligence and machine learning in clinical development: a translational perspective. NPJ Digital Med 2019;2:69. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [16].Rosenthal SA, Hunt D, Sartor AO, et al. A phase 3 trial of 2 years of androgen suppression and radiation therapy with or without adjuvant chemotherapy for high-risk prostate cancer: final results of Radiation Therapy Oncology Group phase 3 randomized trial NRG Oncology RTOG 9902. Int J Radiat Oncol Biol Phys 2015;93:294–302. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [17].Mucherino A, Papajorgji PJ, Pardalos PM. k-Nearest neighbor classification. Data mining in agriculture. New York, NY: Springer New York; 2009. p. 83–106. [Google Scholar]
- [18].Fine JP, Gray RJ. A proportional hazards model for the subdistribution of a competing risk. J Am Stat Assoc 1999;94:496–509. [Google Scholar]
- [19].Gray RJ. A class of K-sample tests for comparing the cumulative incidence of a competing risk. Ann Stat 1988;16:1141–54. [Google Scholar]
- [20].Collins GS, Reitsma JB, Altman DG, Moons KGM, TRIPOD Group. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement. The TRIPOD Group. Circulation 2015;131:211–9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [21].Hendriks RJ, van Oort IM, Schalken JA. Blood-based and urinary prostate cancer biomarkers: a review and comparison of novel biomarkers for detection and treatment decisions. Prostate Cancer Prostatic Dis 2017;20:12–9. [DOI] [PubMed] [Google Scholar]
- [22].Visser WCH, de Jong H, Melchers WJG, Mulders PFA, Schalken JA. Commercialized blood-, urinary- and tissue-based biomarker tests for prostate cancer diagnosis and prognosis. Cancers (Basel) 2020;12:3790. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [23].NCCN. NCCN guidelines prostate cancer. 2022. https://www.nccn.org/professionals/physician_gls/pdf/prostate.pdf
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
All data will be made available as per the NCTN Data Archive rules. The link for the archive is the following: https://nctn-data-archive.nci.nih.gov/.