Unsupervised dense random survival forests identify interpretable patient profiles with heterogeneous treatment benefit

arXiv:2601.01380 · stat.ME · Submitted 2026-01-04 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Unsupervised dense random survival forests identify interpretable patient profiles with heterogeneous treatment benefit".

Paper summary: Jane: So to wrap up this paper, the focus is on using unsupervised dense random survival forests to create interpretable patient profiles that show heterogeneous treatment benefit. The authors are showing that they can identify these subgroups by defining similarity based on treatment-related information, which is a big step toward personalized medicine in cancer care.

Tom: Right. The title itself really sums up the whole point: it's not just finding clusters; it’s making those clusters meaningful by linking them to how the treatment works in biology. They use those dense random forests and that new splitting rule to pinpoint where the treatment effect varies across patients, which is what matters most for clinical decision-making.

Lu: The implication is that this framework provides a systematic way to discover these profiles without needing prior knowledge of exactly what biomarkers will define the response. It’s letting the data speak for itself on how patients respond to therapy, provided you use this specific machine learning structure.

Meng: From a deployment standpoint, it suggests that if we can build models that are inherently interpretable in this way, we might be able to deploy them more confidently because we can see *why* the model is making a certain prediction about a subgroup.

Lalam: Lalam here. The paper applied this framework to two Phase III clinical trials and found that when progression-free survival was the endpoint, patients with wild-type KRAS consistently got benefit from the treatment. When overall survival was used, it showed benefit for patients with wild-type KRAS and also those with an ECOG performance status of zero.

Jane: So, in simple terms, this paper gives clinicians a way to look at patient groups based on their response pattern to treatment rather than just static characteristics. It shows that even when the signal is weak, like with overall survival, this method can reliably find clinically plausible heterogeneity.

Tom: Exactly. It’s about moving from generalized treatments to truly tailored ones by using these dense random survival forests and that novel splitting rule to map out those treatment benefit differences in patients with metastatic colorectal cancer.

Conclusion: Jane: So, we've been looking at how researchers can find hidden patient groups that respond differently to cancer treatment, and this paper is about using these dense random survival forests to do exactly that, identifying profiles with different benefits from therapy.

Tom: Yeah, the title says "unsupervised," which means they aren't telling the computer ahead of time what a group should look like; it finds the patterns on its own based on how patients behave.

Lu: The core idea is using these very dense random survival forests, up to one hundred thousand trees, and a new way of splitting those trees that specifically looks for where the treatment effect changes between different patients.

Meng: From my side, what’s interesting is that they focused on making the results interpretable; it’s not just a black box output but something you can actually look at.

Jane: What does that mean for us as listeners who don't read papers? Basically, it means this method can pull out patient profiles where the benefit from a drug isn't the same for everyone, and they show how to do it without needing pre-defined biomarkers.

Tom: The authors are showing this works on real data from Phase Three trials in colorectal cancer, and they found specific groups that responded better based on things like their tumor genetics or how well they tolerated the treatment.

Lu: The validation is strong because they compared it against scenarios where there was no heterogeneity at all, proving that if there's a difference in treatment benefit, this method can find it.

Meng: From a production standpoint, if we could reliably use this to flag patients who might not respond well to a therapy early on without needing complex genomic sequencing for every single person, that would save time and money.

Jane: It suggests that instead of treating everyone the same way, we can start designing treatments for specific groups right away based on this patient similarity data.

Tom: The implication is that this approach helps move us toward more precise treatment allocation in oncology by identifying those specific patient profiles before treatment even starts.

Lu: So, we've seen how they built the machine learning structure to capture heterogeneity, and now we see it being applied successfully to clinical trial data.

Jane: It really opens the door for thinking about personalized medicine where the "profile" is defined by treatment response rather than just a list of fixed characteristics.

Department of Biostatistics, The University of Texas MD Anderson Cancer Center · Center for Design and Analysis, Amgen

stat.ME

Submitted: 2026-01-04

Updated: 2026-11-17

Code: https://github.com/1996lixingyu1996/DenseRandomForests

Importance score: 83/100

The gist: Unsupervised dense random survival forests identify interpretable patient profiles with heterogeneous treatment benefit> This novel unsupervised machine learning approach based on very dense random…

Key concepts

Dense Random Survival Forests
This is a machine learning method using many decision trees (up to 100,000) to model survival data. It's used here as an unsupervised tool to automatically discover distinct patient groups based on how they respond to treatment, without needing pre-labeled groups.
Novel Splitting Rule
This is a new rule for the random forest that specifically looks for interactions between the treatment effect and other patient characteristics. When this rule is used, it helps the model focus on finding subgroups where the treatment works differently for different patients.
Patient Similarity
Instead of just looking at baseline traits, this method measures how similar two patients are based on their membership in many random forest trees. This similarity is defined by how often they end up in the same final group, which helps define patient profiles based on treatment response.

Terminology

Summary

Unsupervised dense random survival forests identify interpretable patient profiles with heterogeneous treatment benefit> This novel unsupervised machine learning approach based on very dense random survival forests (up to 100,000 trees), equipped with a new splitting rule that explicitly targets treatment-effect heterogeneity, robustly identifies responsive subgroups in precision oncology>

Introduction to the Problem

Colorectal cancer is one of the most common cancer types with more than one million new cases diagnosed annually worldwide Approximately 25% of patients have metastases at diagnosis, and metastases eventually develop in 50% of patients overall Historically, systemic therapy for mCRC relied predominantly on cytotoxic chemotherapy, including fluoropyrimidine-based regimens and combination backbones such as FOLFOX and FOLFIRI The emergence of targeted biologic agents fundamentally altered the therapeutic landscape of mCRC However, it soon became evident that EGFR expression alone was neither necessary nor sufficient to predict therapeutic benefit This paradigm continues to shape contemporary drug development and underscores the need for systematic, biologically informed subgroup investigation in the management of mCRC

Methodology of Similarity Estimation

The method employs a framework where patient similarity is quantified via proximity derived from random forest membership The proximity measure between two observations was originally defined by Breiman as the proportion of trees in which the observations fall into the same terminal node In this work, we apply supervised random forests to characterize patient similarity Pairwise proximity matrices are then derived by calculating the proportion of trees in which two samples occupy the same terminal node The resulting similarity matrix depends on the training parameters of the random forest, such as mtry, nodesize, and others To enhance robustness, we aggregate membership information obtained from multiple random forest models trained under diverse parameter configurations α = α1, · · ·, αk The final similarity matrix is then obtained by fusing membership information across all configurations, yielding what we refer to as a dense random forest

Novel Splitting Rule for Heterogeneity

A novel splitting rule is introduced that jointly accounts for treatment effects while preserving the model’s predictive performance For a given split s, consider a subset of observations indexed by i = 1, · · ·, ns, where ns ≤ n The resulting test statistic (3) from the fitted cox proportional hazards model (2) serves as the splitting rule When ω1 = 0, the model favors identifying covariates that interact with the treatment, thereby facilitating subgroup discovery The optimal split s∗ is defined as the one that maximizes the splitting rule G(s) over all permissible splits

Validation and Clinical Application

Extensive simulations confirm its ability to detect heterogeneous patient responses and distinguish between datasets with and without heterogeneity, while maintaining a stringent Type I error rate of 1% We further validate its performance using Phase III randomized clinical trial datasets, demonstrating significant patient heterogeneity in treatment response based on baseline characteristics In Studies 263 and 309, using PFS as the endpoint, both trials identified a positive treatment effect among patients with wild-type KRAS When OS was used as the endpoint, the estimated treatment benefit was concentrated in patients with wild-type KRAS and ECOG performance status 0 Incorporating both covariates facilitates the characterization of treatment heterogeneity and enhances the clinical interpretability of subgroup analyses

Conclusion on Interpretability

The proposed framework is inherently interpretable, as patient similarity is defined with respect to treatment-related information than solely on baseline covariates Intermediate results of the method, including the t-SNE projections of the fused similarity matrix, provide intuitive visualizations of the patient landscape The final output in the form of clinically meaningful patient profiles further enhances interpretability, making the identified subgroups easier to communicate and potentially actionable in clinical settings The proposed method was applied to two Phase III clinical trials, demonstrating that when PFS was used as the primary endpoint, patients with wild-type KRAS consistently derived benefit from the treatment When OS was considered as the endpoint, treatment benefit was observed among patients with an ECOG performance status of 0, as well as among those with wild-type KRAS The findings suggest that the method can support more precise treatment allocation in clinical practice and provide valuable insights for the design of future clinical trials

Competing Interests

No competing interest is declared The data that support the findings of this study are available from Project Data Sphere The code is available at https://github.com/1996lixingyu1996/DenseRandomForests The authors thank the Amgen Inc. p. 0 for supporting the research

References

Amgen Fda approves first-line use of vectibix (panitumumab) plus folfox Accessed 2025-12-07 Amgen Fda approves vectibix (panitumumab) for use in wild-type ras metastatic colorectal cancer Accessed 2025-12-07 Amgen FDA APPROVES LUMAKRAS® (SOTORASIB) IN COMBINATION WITH VECTIBIX® (PANITUMUMAB) FOR CHEMOREFRACTORY KRAS G12C-MUTATED METASTATIC COLORECTAL CANCER Jan 2025. Accessed 2025-01-17 J C Bendell T S Bekaii-Saab A L Cohn H I Hurwitz M Kozloff H Tezcan N Roach Y Mun S Fish E D Flick et al Treatment patterns and clinical outcomes in patients with metastatic colorectal cancer initially treated with folfox–bevacizumab or folfiri–bevacizumab: results from aries a bevacizumab observational cohort study The oncologist 17(12):1486–1495 2012 L H Biller and D Schrag Diagnosis and treatment of metastatic colorectal cancer a review Jama 325(7): 669–685 2021 C Bokemeyer E Van Cutsem P Rougier F Ciardiello S Heeger M Schlichting I Celik and C-H Köhne Addition of cetuximab to chemotherapy as first-line treatment for kras wild-type metastatic colorectal cancer pooled analysis of the crystal and opus randomised clinical trials European journal of cancer 48(10):1466–1475 2012 F Bray M Laversanne H Sung J Ferlay R L Siegel I Soerjomataram and A Jemal Global cancer statistics 2022 Globocan estimates of incidence and mortality worldwide for 36 cancers in 185 countries CA a cancer journal for clinicians 74(3):229–263 2024 P B Chapman A Hauschild C Robert J B Haanen P Ascierto J Larkin R Dummer C Garbe A Testori M Maio et al Improved survival with vemurafenib in melanoma with braf v600e mutation New England Journal of Medicine 364(26):2507–2516 2011 F Ciardiello and G Tortora Egfr antagonists in cancer treatment New England Journal of Medicine 358(11): 1160–1174 2008 D R Cox Regression models and life-tables Journal of the Royal Statistical Society Series B (Methodological) 34(2):187–202 1972 W Guo Y Ji and D V Catenacci A subgroup cluster-based bayesian adaptive design for precision medicine Biometrics 73(2):367–377 2017 T Hastie R Tibshirani J Friedman et al The elements of statistical learning 2009 L Hu J Ji and F Li Estimating heterogeneous survival treatment effect in observational data using machine learning Statistics in medicine 40(21):4691–4713 2021 N E Hynes and H A Lane Erbb receptors and cancer the complexity of targeted inhibitors Nature Reviews Cancer 5(5):341–354 2005 Y Yarden and M X Sliwkowski Untangling the erbb signalling network Nature reviews Molecular cell biology 2(2):127–137 2001 T Zubair and D Bandyopadhyay Small molecule egfr inhibitors as anti-cancer agents discovery mechanisms of action and opportunities International Journal of Molecular Sciences 24(3):2651 2023 J C Foster J M Taylor and S J Ruberg Subgroup identification from randomized clinical trial data Statistics in medicine 30(24):2867–2880 2011 A Sharafoddini J A Dubin and J Lee Patient similarity in prediction models based on health data a scoping review JMIR medical informatics 5(1):e6730 2017 M Peeters T Price A Cervantes A Sobrero M Ducreux Y Hotko T Andre E Chan F Lordick C Punt et al Final results from a randomized phase 3 study of folfiri±panitumumab for second-line treatment of metastatic colorectal cancer Annals of Oncology 25(1):107–116 2014 J S Rhodes A Cutler and K R Moon Geometry-and accuracy-preserving random forest proximities IEEE Transactions on Pattern Analysis and Machine Intelligence 45(9):10947–10959 2023 P J Rousseeuw Silhouettes a graphical aid to the interpretation and validation of cluster analysis Journal of computational and applied mathematics 20:53–65 1987 S J Ruberg L Chen and Y Wang The mean does not mean as much anymore finding sub-groups for tailored therapeutics Clinical trials 7(5):574–583 2010 D J Sargent B A Conley C Allegra and L Collette Clinical trial designs for predictive marker validation in cancer treatment trials Journal of Clinical Oncology 23(9):2020–27 2005 X Su T Zhou X Yan J Fan and S Yang Interaction trees with censored survival data The international journal of biostatistics 4(1) 2008 Y Kuboki R Yaeger M Fakih J Strickler T Masuishi E Kim C Bestvina C Langer J Krauss S Puri et al Sotorasib with panitumumab in chemotherapy-refractory kras g12c-mutated colorectal cancer a phase 1b trial Nature medicine 30(1):265–270 2024 J Lee D M Maslove and J A Dubin Personalized mortality prediction driven by electronic medical data and a patient similarity metric PloS one 10(5):e0127428 2015 M Panahiazar V Taslimitehrani N L Pereira and J Pathak Using ehrs for heart failure therapy recommendation using multidimensional patient similarity analytics Studies in health technology and informatics 210:369 2015 E Parimbelli S Marini L Sacchi and R Bellazzi Patient similarity for precision medicine A systematic review Journal of biomedical informatics 83:87–96 2018 M Peeters T Price A Cervantes A Sobrero M Ducreux Y Hotko T Andre E Chan F Lordick C Punt et al Final results from a randomized phase 3 study of folfiri±panitumumab for second-line treatment of metastatic colorectal cancer Annals of Oncology 25(1):107–116 2014 J S Rhodes A Cutler and K R Moon Geometry-and accuracy-preserving random forest proximities IEEE Transactions on Pattern Analysis and Machine Intelligence 45(9):10947–10959 2023 S J Ruberg L Chen and Y Wang The mean does not mean as much anymore finding sub-groups for tailored therapeutics Clinical trials 7(5):574–583 2010 D J Sargent B A Conley C Allegra and L Collette Clinical trial designs for predictive marker validation in cancer treatment trials Journal of Clinical Oncology 23(9):2020–27 2005 A Sharafoddini J A Dubin and J Lee Patient similarity in prediction models based on health data a scoping review JMIR medical informatics 5(1):e6730 2017 S Siena A Sartore-Bianchi F Di Nicolantonio J Balfour and A Bardelli Biomarkers predicting clinical outcome of epidermal growth factor receptor–targeted therapy in metastatic colorectal cancer Journal of the National Cancer Institute 101(19):1308–1324 2009 S Sivaganesan P W Laud and P Müller A bayesian subgroup analysis with a zero-enriched polya urn scheme Statistics in medicine 30(4):312–323 2011

Simulation Results Comparison

The simulation study showed that the proposed method accurately identified true treatment-effect heterogeneity when it was present and correctly detected the absence of heterogeneity in the global and null scenarios In Scenario 1, the first image in the top row depicted the ground truth, with the true region located in the upper right corner The second image presented the simulation result produced by our method, illustrating that the identified subgroups closely align with the true region In contrast, the third image, generated using Kmeans clustering, showed a notable deviation from the ground truth Overall, the simulation study showed that the proposed method accurately identified true treatment-effect heterogeneity when it was present and correctly detected the absence of heterogeneity in the global and null scenarios

Clinical Trial Findings

In Studies 263 and 309, using PFS as the endpoint, both trials identified a positive treatment effect among patients with wild-type KRAS When OS was used as the endpoint, the estimated treatment benefit was concentrated in patients with wild-type KRAS and ECOG performance status 0 Compared with the PFS-based analysis, the OS-based results indicate effect modification by ECOG status KRAS mutation status and ECOG performance status were both clinically and biologically meaningful covariates in colorectal cancer, making them reasonable and well-justified factors for subgroup analyses These findings indicate that, even under a low signal-to-noise ratio for OS, the proposed method reliably recovered clinically plausible treatment-effect heterogeneity

Discussion on Method Advantages

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved patient subgroup identification using Dense Random Forests can identify interpretable patient profiles with heterogeneous treatment benefit by explicitly targeting treatment-effect heterogeneity through a novel splitting rule that assesses quality based on a Cox proportional hazards model.

  2. The system can perform similarity-based personalized prediction by constructing a proximity matrix where proximity serves as an empirical estimator of pairwise similarity, which is then used for clustering, allowing it to capture complex, non-Euclidean cluster structures compared to traditional distance-based methods like K-means.

  3. The improved AI system can distinguish between true treatment effect heterogeneity and random fluctuations by employing a custom metric that selects the most representative patient profile among results obtained under different clustering specifications, controlling the Type I error rate at 1% when calibrated under null scenarios.

  4. The framework enables more robust subgroup discovery through an ensemble proximity matrix, where this ensemblebased representation substantially mitigates the sensitivity of subgroup identification to tuning parameters of individual random forest models.

  5. The system can generate clinically meaningful patient profiles by utilizing decision trees which stratify patients into subgroups on the basis of their attributes while facilitating flexibility to handle mixed types of data, making the identified subgroups easier to communicate and potentially actionable in clinical settings.

Related papers