Prostate Cancer: Early Detection and Assessing Clinical Risk Using Deep Machine Learning of High Dimensional...

Area of Science:

• Immunology and Cancer Diagnostics
• Artificial Intelligence in Medicine
• Prostate Cancer Research

Background:

• Early detection of prostate cancer (PCa) and risk stratification are crucial for effective treatment, yet current methods like biopsies can be invasive.
• Distinguishing high-risk PCa from low- or intermediate-risk disease without invasive procedures remains a significant clinical challenge.
• Previous research identified specific T and B cell phenotypic features capable of differentiating benign prostate disease from PCa in men with Prostate-Specific Antigen (PSA) levels < 20 ng/ml.

Purpose of the Study:

• To evaluate if previously identified T and B cell phenotypic features can detect PCa presence and clinical risk in a broader cohort of men with elevated PSA levels (3-2617 ng/ml).
• To develop and validate AI models, specifically bidirectional Long Short-Term Memory Deep Neural Networks (biLSTM), for PCa detection and risk prediction.
• To assess the performance of AI models combining immunophenotypic features with clinical data (Age, PSA) for improved diagnostic accuracy.

Main Methods:

• Immune profiling of peripheral blood from 130 asymptomatic men with elevated PSA levels using multiparametric whole blood flow cytometry.
• Development of two biLSTM models: one for PCa presence detection (combining T/B cell phenotypes and Age) and another for high-risk PCa prediction (combining phenotypes and PSA).
• Biopsy-based diagnosis confirmed 42 cases of benign prostate disease and 88 cases of PCa among the study participants.

Main Results:

• The PCa detection model achieved high performance metrics, including an accuracy of 86.79% (± 0.10), sensitivity of 82.78% (± 0.15), and specificity of 95.83% (± 0.11) on the test set.
• The risk prediction model demonstrated strong performance with an accuracy of 94.90% (± 6.29) and specificity of 96.11 (± 0.00) for identifying high-risk PCa.
• Combining flow cytometry (FC) data with PSA levels resulted in a lower false positive rate (ORP-FPR) for PCa detection compared to PSA alone.

Conclusions:

• AI-driven analysis of peripheral blood T and B cell phenotypes can effectively distinguish between benign prostate disease and PCa.
• The developed AI models accurately predict the clinical risk of prostate cancer in asymptomatic men with elevated PSA levels.
• This approach offers a promising, non-invasive strategy for PCa diagnosis and risk stratification, potentially reducing the need for unnecessary biopsies.