TL;DR
This study explores unsupervised clustering as a method to stratify non-small-cell lung cancer (NSCLC) patients with brain metastases (BM) into prognostic subgroups. The findings suggest that unsupervised methods like Hierarchical Cluster Analysis (HCA) and Two-Step Clustering (TSC) can identify distinct patient groups with significantly different survival outcomes, performing comparably to the established Diagnosis-Specific Graded Prognostic Assessment (DS-GPA).
What Problem Does This Paper Address?
Predicting the prognosis for patients with non-small-cell lung cancer (NSCLC) that has metastasized to the brain (BM) is crucial for guiding treatment decisions and managing patient expectations. While established prognostic scoring systems like the Graded Prognostic Assessment (GPA) and its variants (e.g., DS-GPA) exist, they often rely on predefined categories and may not fully capture the complex interactions of clinical variables. The authors highlight the increasing reliance on complex, computationally intensive AI systems and propose that simpler, yet effective, unsupervised approaches could offer a valuable alternative for patient stratification, potentially reducing the burden on clinicians and uncovering novel prognostic patterns.
Methods
The researchers employed two unsupervised clustering techniques:
- Two-Step Clustering (TSC): This method uses a pre-clustering stage to build a clustering feature (CF) tree based on a log-likelihood distance measure, suitable for mixed data types. A subsequent hierarchical clustering stage then derives the final clusters, with the number of clusters automatically determined by the Bayesian Information Criterion (BIC).
- Hierarchical Cluster Analysis (HCA): This approach uses Euclidean distance after standardizing continuous variables and numerically encoding categorical ones. Ward's linkage method was chosen for its tendency to create compact clusters. The number of clusters was determined by visual inspection of the dendrogram.
Both methods were applied to baseline clinical data from NSCLC patients with BM. The prognostic performance of these clusters was then compared to the established DS-GPA classification using Kaplan-Meier survival analysis and log-rank tests. The discriminative ability of each classification method was quantified using the concordance index (C-index).
Datasets
The study utilized a retrospective dataset of 95 NSCLC patients with brain metastases (BM) treated at two oncology centers between 2015 and 2021. The data included demographic features, prognostic variables (age, Karnofsky performance status (KPS), extracranial metastasis (ECM) presence, number of brain metastases), and overall survival (OS) times. Molecular subtypes (EGFR and ALK statuses) were excluded from the clustering analysis as all patients had negative or unknown statuses.
Key Insights and Ideas
- Unsupervised learning can uncover clinically relevant prognostic subgroups: The study demonstrates that algorithms like TSC and HCA can identify distinct patient groups with significantly different survival outcomes, even without prior knowledge of these outcomes.
- HCA shows promising discriminatory power: Hierarchical Cluster Analysis, in particular, achieved the highest C-index, suggesting it may be a powerful tool for stratifying patients with BM.
- Unsupervised methods offer a data-driven perspective: Unlike rule-based indices, unsupervised clustering can capture complex, non-linear interactions among variables, potentially leading to more nuanced prognostic assessments.
- Potential for personalized treatment: The identified subgroups could inform treatment decisions, guiding more aggressive therapies for favorable groups and palliative care for those with poorer prognoses.
- Adaptability to new data: Unsupervised frameworks are flexible and can readily incorporate new data modalities (e.g., radiomic or multi-omic features) as they become available.
Key Results
- The study included 95 NSCLC patients with BM, with a mean age of 61.8 years and 77.9% being male.
- The median overall survival (OS) for the cohort was 8 months.
- Both TSC and HCA identified four distinct clusters, and these clusters showed statistically significant differences in OS (p < 0.001 for both).
- The DS-GPA classification also significantly divided patients into prognostic classes (p < 0.001).
- Discriminatory Power (C-index):
- Hierarchical Cluster Analysis (HCA): 0.721
- Diagnosis-Specific Graded Prognostic Assessment (DS-GPA): 0.709
- Two-Step Clustering (TSC): 0.650
- KPS was a significant differentiator across all clustering methods and the DS-GPA.
- Extracranial metastasis (ECM) was exclusively found in one cluster (Cluster 3 in both TSC and HCA) and showed a significant difference.
- The number of brain metastases also varied significantly between clusters.
Limitations and Future Work
- Retrospective design: The study's retrospective nature may introduce biases.
- Small sample size: The relatively small sample size (n=95) limits the ability to perform external or internal cross-validation. Future multi-center studies are needed for confirmation.
- Exclusion of molecular subtypes: EGFR and ALK mutation statuses were excluded because they were uniformly negative or unknown, meaning the prognostic classifications did not account for the impact of molecular and histological subtypes. This limits generalizability to molecularly stratified populations.
- Not suitable for clinical application yet: Due to these limitations, the current model is not yet ready for direct clinical application.
The authors suggest that future work should focus on prospective validation and methodological refinement to better understand the scope and limitations of unsupervised clustering for prognostic modeling. They also propose collecting multi-omic data to enhance the performance of unsupervised learning by leveraging interactions among a wider range of features.
Summary updated