Prostate cancer is the most common cancer among American men, with nearly 300,000 new diagnoses and 35,000 deaths in 2024. Men of African ancestry face disproportionately higher rates of both diagnosis and mortality compared to men of European ancestry.
The current gold standard for staging prostate cancer is the American Joint Committee on Cancer (AJCC) Tumor, Lymph Node, and Metastasis (TNM) Staging System. It uses five clinical factors: tumor extent (T), lymph node spread (N), distant metastasis (M), PSA levels (P), and Grade Group/Gleason score (G).
Despite its widespread use, the AJCC TNM system has meaningful limitations. It omits potentially important variables like age and race, shows weak separation between certain stages (such as Stage I and Stage IIA), and has limited accuracy in predicting disease progression and recurrence.
A patient's prognosis changes dramatically by stage: five-year survival is nearly 100% for localized prostate cancer but drops to only 33% for patients with distant metastasis. More accurate staging tools are therefore critical for guiding treatment and improving outcomes.
Researchers applied the Ensemble Algorithm for Clustering Cancer Data (EACCD), an unsupervised machine learning approach developed specifically for cancer survival data. The algorithm works in three main steps: calculating initial dissimilarities between patient groups, performing ensemble learning, and conducting hierarchical clustering.
The key idea behind EACCD is to group patients by how similar their survival outcomes are, rather than by arbitrary clinical thresholds. The algorithm measures survival differences between patient combinations using the Gehan-Wilcoxon test statistic and then clusters them using minimax linkage hierarchical clustering.
Data came from the SEER (Surveillance, Epidemiology, and End Results) database maintained by the National Cancer Institute. The study used 161,212 patients diagnosed between 2010 and 2014 as training data, and 29,161 patients diagnosed in 2015 as a separate validation dataset.
Two versions of the model were built: a five-variable model using the same T, N, M, PSA, and Grade Group variables as AJCC, and an extended seven-variable model that added age and race. Any patient combination with fewer than 25 patients was excluded to maintain statistical reliability.
The seven prognostic variables used in the extended model include: tumor stage (T), ranging from non-palpable tumors (T1) to those spreading beyond the prostate (T4); lymph node involvement (N); and distant metastasis (M), which includes spread to lymph nodes, bones, or other organs.
PSA (prostate-specific antigen) levels were categorized into three groups: P1 (less than 10 ng/mL), P2 (10 to 20 ng/mL), and P3 (20 ng/mL or higher). Grade Group ranges from G1 (well-formed glands, Gleason score 6 or below) to G5 (lacking gland formation, Gleason score 9-10).
Age was split into two groups: under 70 (A0) and 70 or older (A1). Race was categorized as White/European ancestry (R1), Black/African ancestry (R2), or Other (R3). Both of these variables have established prognostic significance but are not included in the current AJCC staging system.
The training dataset was broadly representative: 50% of patients had T1 tumors, 98% had no lymph node involvement, 44% were Grade Group 1, and 79% were of White ancestry. This demographic distribution closely matched the broader SEER prostate cancer population.
The performance of the models was measured using Harrell's concordance index (C-index), where a score of 1.0 represents perfect prediction and 0.5 represents random chance. The AJCC 9th edition staging system achieved a C-index of 0.7676.
The five-variable EACCD model clustered patients into 10 distinct prognostic groups and achieved a C-index of 0.8293, significantly outperforming the AJCC system. The survival curves for all 10 groups were clearly separated, with statistically significant differences between adjacent groups (log-rank test p less than 0.05).
The seven-variable EACCD model, which added age and race, further improved performance to a C-index of 0.8504 and generated 13 distinct prognostic groups. This model demonstrated that including demographic factors alongside clinical variables yields meaningfully better survival prediction.
In contrast, the AJCC survival curves showed weak separation between several stages. Stage I closely overlapped with Stage IIA, and Stage IIC overlapped with Stage IIIA, highlighting the practical limitations of the current standard staging system.
To confirm the models were not simply overfitting to the training data, both EACCD systems were tested on an independent validation dataset of 29,161 patients diagnosed in 2015. Consistency between training and validation performance is a key indicator of a reliable prognostic model.
The five-variable model maintained strong performance on the validation data, achieving a C-index of 0.8437 -- actually slightly higher than the training C-index of 0.8293. The seven-variable model similarly held up, with a validation C-index of 0.8585 compared to 0.8504 in training.
Survival curve separation in the validation dataset was also well-maintained. For the five-variable system, 7 out of 9 adjacent group comparisons were statistically significant; for the seven-variable system, 9 out of 12 were significant. In the remaining cases, hazard ratios still pointed in the expected direction of risk ordering.
This consistency across independent datasets strongly supports the reproducibility and robustness of the EACCD prognostic systems for potential clinical use.
A key strength of unsupervised clustering algorithms like EACCD is their ability to discover novel patient subgroups without being constrained by existing knowledge. Traditional supervised methods are designed to reinforce known patterns, while EACCD can reveal hidden structures in complex datasets.
Compared to a previous effort by Sun et al., which modified the AJCC system using a nomogram and achieved a C-index of 0.789 on a dataset of 16,803 patients, the EACCD approach achieved higher accuracy (C-index up to 0.8504) on a much larger dataset of over 161,000 patients. Nomograms also rely on the proportional hazards assumption, which is frequently violated in real-world data.
Other supervised machine learning studies -- including those using XGBoost to predict biochemical recurrence or pathological stage -- have shown promise, but they predict specific endpoints rather than providing a comprehensive prognostic staging framework. EACCD offers a more holistic approach to patient risk stratification.
The EACCD system is also adaptable: it can be updated as new prognostic factors are discovered, such as circulating tumor DNA (ctDNA) and lactate dehydrogenase levels. This flexibility makes it a strong candidate for integration into future clinical decision support tools.
Using the EACCD system in practice is straightforward. A clinician determines the patient's values for each relevant variable (T, N, M, PSA, Grade Group, and optionally age and race), assigns the patient to a prognostic group based on those values, and then uses the corresponding survival curve to communicate prognosis and guide treatment decisions.
This algorithmic approach moves beyond committee-based, expert-opinion staging toward a data-driven framework grounded in actual survival outcomes from hundreds of thousands of patients. The system can help identify patients who might otherwise be miscategorized under the current AJCC staging.
The inclusion of race as a prognostic variable is particularly notable. African American men have higher prostate cancer mortality rates, and ignoring this difference in a staging system may lead to under-recognition of risk in this population. EACCD explicitly accounts for this disparity.
With further validation in prospective cohorts, the EACCD framework could support precision oncology -- tailoring treatment intensity and follow-up schedules to each patient's actual risk level rather than a broad stage category.
The EACCD machine learning approach successfully created prognostic systems for primary prostate cancer that outperform the current AJCC TNM staging standard in predictive accuracy. Both the five-variable and seven-variable models produced well-separated survival groups and higher C-index scores.
The extended seven-variable model, which incorporates age and race, demonstrated the added prognostic value of these demographic factors. This supports the argument that modern staging systems should move beyond purely anatomical and pathological variables.
Validation results confirmed that the EACCD models are reproducible and reliable, a critical requirement for clinical adoption. The consistent performance on independent data suggests that these models reflect real biological differences in prostate cancer outcomes rather than statistical noise.
Future work should focus on validating EACCD in independent prospective cohorts, incorporating additional biomarkers as they become available, and developing a user-friendly clinical decision support tool that clinicians can deploy at the point of care to improve prostate cancer management.