Developing machine learning algorithms for dynamic estimation of progression during active surveillance for prostate cancer

Changhee Lee, Alexander Light, Evgeny S. Saveliev, Mihaela van der Schaar, Vincent J. GnanapragasamNpj Digit. Medicine 2022 · OpenReview

TL;DR

This paper introduces a novel machine learning framework, Dynamic-DeepHit-Lite (DDHL), that dynamically estimates prostate cancer progression risk using longitudinal patient data. By integrating real-time updates from clinical monitoring, DDHL significantly improves predictive accuracy over traditional methods, offering a more personalized approach to active surveillance.

What Problem Does This Paper Address?

Active Surveillance (AS) is a common management strategy for early-stage prostate cancer, involving regular monitoring rather than immediate treatment. However, a key challenge in AS is accurately predicting the risk of disease progression over time, especially as new clinical data (like PSA levels, MRI scans, and biopsy results) becomes available. Existing models often rely solely on baseline patient characteristics, failing to incorporate the "live" updates that are crucial for timely intervention decisions. This gap leaves clinicians without a robust tool to dynamically adjust risk assessments and personalize follow-up strategies for individual patients.

Methods

The researchers developed two core machine learning components:

  • Dynamic-DeepHit-Lite (DDHL): This is a deep learning-based survival model designed for individualised longitudinal time-to-event analysis. DDHL uses a recurrent neural network (RNN) to learn data-driven distributions of time-to-event outcomes, effectively incorporating sequential clinical data (PSA, MRI, biopsy) collected over time. It's an adaptation of the Dynamic-DeepHit model, refined for better performance with smaller datasets and improved parameter efficiency.
  • Actor-Critic for Temporal Predictive Clustering (AC-TPC): This reinforcement learning approach is used to discover distinct groups (clusters) of patients who share similar time-to-event trajectories. AC-TPC leverages the DDHL model's outputs to partition patients into these clusters, providing a more interpretable, outcome-oriented view of progression risk.

These methods were trained and validated using a combination of static baseline covariates and longitudinal follow-up data.

Datasets

The study utilized a de-identified dataset of 585 men enrolled in a structured Active Surveillance program. This dataset included:

  • Baseline characteristics: Age, ethnicity, family history, risk assignment (Cambridge Prognostic Group), MRI stage, and Likert score at entry.
  • Longitudinal follow-up data: Serial PSA tests, multi-parametric MRI measurements (including PRECISE scoring), and repeat biopsy information.

The median follow-up period was 4.4 years. The data was split into training and testing sets for model evaluation.

Key Insights and Ideas

  • The power of dynamic updates: The most significant insight is that incorporating longitudinal, "live" data dramatically improves the prediction of prostate cancer progression compared to models relying solely on baseline information. This highlights the limitations of static risk assessment in a dynamic disease context.
  • Temporal clustering for clinical interpretability: Beyond individual risk scores, identifying distinct temporal clusters of patients with similar progression trajectories offers a more intuitive way for clinicians to understand and manage risk. This moves beyond a single risk number to a more nuanced understanding of a patient's likely disease path.
  • DDHL's adaptability: The DDHL model's ability to learn from sequential data and adapt its predictions as new information becomes available is crucial. This allows for continuous risk refinement, a key requirement for effective active surveillance.
  • PRECISE scoring integration: The study's inclusion of the PRECISE scoring system for MRI interpretation is notable, as it's a relatively new metric and its integration into ML models for AS is a step towards leveraging advanced imaging data.

Key Results

  • Improved Discrimination: The DDHL model showed superior predictive performance compared to traditional Cox regression and landmarking methods, especially when incorporating follow-up data.
  • With 3 years of data collection and 3 years of follow-up evaluation, DDHL achieved a C-index of 0.79 (±0.11), outperforming landmarking Cox (0.70 ±0.15) and baseline Cox (0.67 ±0.09).
  • At 5 years follow-up, DDHL's C-indices were 0.82 (±0.08), compared to 0.75 (±0.08) for landmarking Cox and 0.73 (±0.09) for baseline Cox.
  • Temporal Clusters: The AC-TPC method identified 4 distinct temporal clusters with varying progression risks.
  • The lowest risk cluster had a negligible progression risk (<5% by 8 years).
  • The highest risk cluster had a >50% risk of progression by 5 years.
  • Model Calibration: Model calibration was reported as good across tested models, particularly at the 3-year follow-up horizon.
  • Comparison to other models: The study's 4-cluster model demonstrated higher discriminative power than a recently reported dynamic risk prediction model (Canary PASS group), especially over longer time horizons.

Limitations and Future Work

  • Cohort Size and Single Center: The study is based on a relatively small, single-center cohort, which limits generalizability.
  • Follow-up Duration: The median follow-up of less than 5 years, while clinically useful, could be extended.
  • Missing Data: Serial MRI data had a significant amount of missing values, and the impact of this missingness on model performance is unclear.
  • Endpoint Definition: The chosen endpoint (progression to ≥CPG3 disease) is not universally accepted, and external validation with more common endpoints is needed.
  • Future Work: The authors suggest external validation with larger cohorts and more standardized endpoints. They also propose incorporating lifestyle factors and comorbidities into the model to further enhance its predictive power. The development of a clinical demonstrator webtool highlights the potential for real-time clinical application.

Summary updated

Abstract

Active Surveillance (AS) for prostate cancer is a management option that continually monitors early disease and considers intervention if progression occurs. A robust method to incorporate “live” updates of progression risk during follow-up has hitherto been lacking. To address this, we developed a deep learning-based individualised longitudinal survival model using Dynamic-DeepHit-Lite (DDHL) that learns data-driven distribution of time-to-event outcomes. Further refining outputs, we used a reinforcement learning approach (Actor-Critic) for temporal predictive clustering (AC-TPC) to discover groups with similar time-to-event outcomes to support clinical utility. We applied these methods to data from 585 men on AS with longitudinal and comprehensive follow-up (median 4.4 years). Time-dependent C-indices and Brier scores were calculated and compared to Cox regression and landmarking methods. Both Cox and DDHL models including only baseline variables showed comparable C-indices but the DDHL model performance improved with additional follow-up data. With 3 years of data collection and 3 years follow-up the DDHL model had a C-index of 0.79 (±0.11) compared to 0.70 (±0.15) for landmarking Cox and 0.67 (±0.09) for baseline Cox only. Model calibration was good across all models tested. The AC-TPC method further discovered 4 distinct outcome-related temporal clusters with distinct progression trajectories. Those in the lowest risk cluster had negligible progression risk while those in the highest cluster had a 50% risk of progression by 5 years. In summary, we report a novel machine learning approach to inform personalised follow-up during active surveillance which improves predictive power with increasing data input over time.
BibTeX
@inproceedings{lee2022developing,
  author = {Changhee Lee and Alexander Light and Evgeny S. Saveliev and Mihaela van der Schaar and Vincent J. Gnanapragasam},
  title = {{Developing machine learning algorithms for dynamic estimation of progression during active surveillance for prostate cancer}},
  booktitle = {Npj Digit. Medicine},
  year = {2022},
  url = {https://openreview.net/forum?id=_KCoVjaVqV},
}
Figure