ΑΙhub.org

**By Chirag Nagpal and Willa Potosnak**

Real-world decision-making often requires reasoning about **when** an event will occur. The overarching goal of such reasoning is to help aid decision-making for optimal triage and subsequent intervention. Such problems involving estimation of Times-to-an-Event frequently arise across multiple application areas, including,

**Healthcare and Bio-informatics**: More commonly known as ‘*Survival Analysis*‘ involves prognostication of an adverse physiological event like a stroke, the onset of cancer, re-hospitalization, and mortality. Time-to-event or survival analysis can be used to proactively mitigate adverse outcomes and extend the longevity of patients.**Internet Marketing and e-commerce**: Models employed for estimating customer churn and retention in large commercial organizations are essentially time-to-event regression models and help determine best practices to maximize customer retention.**Predictive Maintenance**: Reliability engineering and systems safety research involves the use of remaining useful life prediction models to help extend the longevity of machinery and equipment by proactive part and component replacement.**Finance and Actuarial and Sciences**: Time-to-Event models are ubiquitous in the estimation of optimal financial strategies for setting insurance premiums, as well as estimating credit defaulting behavior.

Discretizing time-to-event outcomes to predict** if** an event will occur is a common approach in standard machine learning. However, this neglects temporal context, which could result in models that misestimate and lead to poorer generalization.

In our **recent Machine Learning for Healthcare ’22 paper**, we present – a comprehensive Python code repository of user-friendly, machine learning tools for working with censored time-to-event data. This package includes an exclusive suite of workflows for a range of tasks from data pre-processing and regression modeling to model evaluation.

`auton-survival`

`auton-survival`

`scikit-learn`

Time-to-Event or Survival regression can be used to estimate the conditional probability of an event occurring within a specified time period or event-horizon. A time-to-event estimation problem thus reduces to estimating the conditional distribution of survival:

Note that is a set of covariates, and refers to the distribution of the censored survival time where is the distribution of the true time-to-event and is the distribution of the censoring time. Assuming conditional independence between and (ie., ) allows identification of the distribution of .

Survival regression naturally allows accounting for censored data. In the case of survival regression, the likelihood under censoring is given as

.

Here are the covariates, is the event or censoring time and is a binary indicator denoting if the individual was censored. For the censored individuals, the likelihood corresponds to the probability that the event takes place beyond the time horizon, also known as the ‘*survival function*‘.

Broadly, the popular approaches for learning estimators of survival in the presence of censoring can be categorized into:

**Parametric:**Assume that time-to-event distribution adheres to a known parametric distribution, such as Weibull or Log-Normal.**Non-Parametric**: Involve learning kernels or similarity functions of the input covariates followed by a non-parametric (Kaplan-Meier or Nelson-Aalen) estimation of the survival rate weighted with the learned kernel.**Semi-Parametric**: As with Cox Proportional Hazards models, feature interactions are learned through a parametric model followed by a non-parametric estimation of the base survival (hazard) rate.

Complex multimodal data often observed in healthcare and other applications, bring a multitude of challenges to traditional machine learning. allows a simple interface to use deep neural networks and representation learning to model such complex data.

`auton-survival`

includes extensions to the standard Cox Proportional Hazards (CPH) (Cox, 1972) involving deep representation learning (Faraggi and Simon, 1995; Katzman et al., 2018) as well as latent variable survival regression models,

`auton-survival`

**The SurvivalModel Class**

The package provides a convenient ** SurvivalModel** class that enables rapid experimentation via a consistent API that wraps multiple alternative regression estimators. In addition to the models mentioned above, the

`SurvivalModel`

Hyperparameter tuning for model selection can be streamlined with the ** SurvivalRegressionCV** class to apply cross-validation over a user-specified hyperparameter grid.

**Time-Varying Survival Regression**

Real-world data often consists of multiple time-dependent observations per individual or time-varying covariates. is equipped to handle time-varying covariates for survival analysis with auto-regressive deep learning models that allow learning temporal dependencies when estimating time-to-event outcomes. Implementations of time-varying DSM and Deep Cox Proportional Hazards model involve the use of RNNs, LSTMs, or GRUs (Chung et al., 2014; Hochreiter and Schmidhuber, 1997) for time-varying survival regression as shown in

`auton-survival`

**Counterfactual Estimators of Survival**

Decision support often requires reasoning about * ‘what if’* scenarios regarding the effect of different treatments on outcomes. In observational settings, outcomes and treatment assignments may share common causes. Adjusting for such confounding factors is crucial when performing causal inference.

`auton-survival`

where is just an estimate of the conditional expectation of survival learnt on the population under intervention* *.

Consider the data from the large SEER Cancer Incidence registry (Ries et al., 1975). When stratified by region (**Figure 4a**), there is an apparent disparity in survival rates. We demonstrate the use of counterfactual regression to provide insight into whether these discrepancies can be attributed to the geographic region or other socio-economic or physiological confounding factors that affect both belonging to these regions and the outcomes. To adjust estimates of survival with counterfactual estimation, we train two separate Deep Cox models on data from Greater California and Louisiana as counterfactual regressors. The fitted regressors are then applied to estimate the survival curves for each instance, which are then averaged over treatment groups to compute the domain-specific survival rate. **Figure 4b** presents the counterfactual survival rates compared with the survival rates obtained from a Kaplan-Meier estimator. The Kaplan-Meier estimator does not adjust for confounding factors and overestimates treatment effect, as evidenced by the extent that survival rates differ between regions. Alternatively, counterfactual regression adjusts for confounding factors and predicts more similar survival rates between regions.

Figure 5:

Survival rates differ across groups of individuals with heterogeneous characteristics. Identifying groups of patients with similar survival rates can be used to derive insight into practices and interventions that can help improve longevity for such groups. While domain knowledge can help identify such subgroups, in practice there could be potentially complex, non-linear feature interactions that determine assignment to subgroups, making identification difficult. In , we refer to this group identification and survival assessment as

`auton-survival`

Our package offers multiple approaches to phenotyping that involve either the use of specific domain knowledge, as in the case of the intersectional phenotyper, or a completely unsupervised approach that clusters subjects based on the observed covariates. Additionally, also offers phenotypers that explicitly involve supervision in the form of the observed outcomes and counterfactuals inform the learned phenotypes to better stratify the data. Directed Acyclic Graphical representations of probabilistic phenotypers in

`auton-survival`

`auton-survival`

**Intersectional Phenotyping**: Recovers groups, or phenotypes, of individuals over exhaustive combinations of user-specified categorical and numerical features.Identifies groups of individuals based on structured similarity in the feature space by first performing dimensionality reduction of the input covariates, followed by clustering. The estimated probability of an individual belonging to a latent group is computed as the distance to the cluster normalized by the sum of distances to other clusters.**Unsupervised Phenotyping**:Identifies latent groups of individuals with similar survival outcomes conditioned on outcomes. This approach can be performed as a direct consequence of training the DSM and DCM latent variable survival estimators.**Supervised Phenotyping**:Identifies groups of individuals that demonstrate enhanced or diminished treatment effects (Chirag et al., 2022).**Counterfactual Phenotyping**:

**Figure 6** presents the Kaplan-Meier survival curves of the phenogroups extracted from SUPPORT (Knaus et al., 1995) using the unsupervised and supervised phenotyping. The intersecting survival curves suggest the phenotypers’ ability to recover phenogroups that do not strictly adhere to assumptions of Proportional Hazards. From **Figure 7**, it can be inferred that supervised phenotyping extracts phenogroups with higher discriminative power as indicated by the contrasting Kaplan-Meier estimates of phenogroup level survival.

offers additional tools to analyze the effect of an intervention on outcomes by computing propensity-adjusted treatment effects in terms of the following metrics through bootstrap resampling of the dataset with replacement:

`auton-survival`

**Hazard Ratio**: Assuming the proportional hazards assumptions holds, the treatment effect can be measured as the ratio of hazard rates between the treatment and control arms.**Time at Risk (TaR)****(Figure 7a)**: The treatment effect can be measured as the difference in time-to-event at a specified level of risk.**Risk at Time (Figure 7b)**: The treatment effect can be measured as the difference in risk at a specified time horizon.**Restricted Mean Survival Time (RMST) (Figure 7c):**The treatment effect can be measured as the difference in the expected (or mean) time-to-event conditioned on a specified time horizon.

Propensity-adjustment allows an alternative approach to estimate treatment effects of potential confounders that influence both treatment assignment and the outcome. Not adjusting for treatment propensity could result in misestimations of treatment effects.

allows adjusting for treatment propensity with computation of treatment effects bootstrapped with sample weights. When the specified sample weights are propensity scores, such as obtained from a classification model, the bootstrapped distribution treatment effect converges to the Inverse Propensity of Treatment Weighting (IPTW) Thompson-Horvitz estimate of the population Average Treatment Effect.

`auton-survival`

In a second analysis of the effect of geographical region on breast cancer mortality using data from the SEER cancer registry (Ries et al., 1975), we compare treatment effects before and after adjusting for confounding factors by inverse propensity weighting. Similar to the previous analysis with counterfactual regression, we consider the regions of “Greater California” and “Louisiana” as the binary “treatment” in question. To adjust treatment effects for confounding factors, we first trained a logistic regression with an penalty by regressing the geographical region on the set of confounding variables. The estimated propensity scores are then employed as sampling weights for the treatment effects in terms of hazard ratios, restricted mean survival time (RMST), and risk difference as in **Figure 8**. Adjusting for region propensity noticeably mitigates differences in treatment effects, indicating that mortality due to breast cancer is likely explained by confounding socio-economic and physiological factors rather than solely the geographic region.

We present , an open-source Python package encapsulating multiple pipelines to work with censored time-to-event data. Such data is ubiquitous in many fields, including healthcare and the maintenance of equipment. Through continuous collaboration with the machine learning for healthcare community, we aim to better aid machine learning research in efforts to create a robust, comprehensive repository of rigorous tools for reproducible analysis of censored time-to-event data.

`auton-survival`

**Chirag Nagpal**PhD Candidate, Auton Lab

@nagpalchirag

cs.cmu.edu/~chiragn

**Willa Potosnak**Research Intern and

Incoming PhD Student, Auton Lab

potosnakw.github.io

**[1] **Nagpal, C., Potosnak, W. and Dubrawski, A., 2022. auton-survival: an Open-Source Package for Regression, Counterfactual Estimation, Evaluation and Phenotyping with Censored Time-to-Event Data. *arXiv preprint arXiv:2204.07276*.

**[2]** Fabian Pedregosa, Ga¨el Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. Scikit-learn: Machine learning in python. the Journal of machine Learning research, 12: 2825–2830, 2011.

**[3] **D. R. Cox. Regression models and life-tables. Journal of the Royal Statistical Society. Series B (Methodological), 34(2):187–220, 1972.

**[4]** David Faraggi and Richard Simon. A neural network model for survival data. Statistics in medicine, 14(1):73–82, 1995.

**[5]** Jared L Katzman, Uri Shaham, Alexander Cloninger, Jonathan Bates, Tingting Jiang, and Yuval Kluger. Deepsurv: personalized treatment recommender system using a cox proportional hazards deep neural network. BMC medical research methodology, 18(1): 1–12, 2018.

**[6]** Nagpal, C., Li, X. and Dubrawski, A., 2021a. Deep survival machines: Fully parametric survival regression and representation learning for censored data with competing risks. *IEEE Journal of Biomedical and Health Informatics*, *25*(8), pp.3163-3175.

**[7]** Nagpal, C., Yadlowsky, S., Rostamzadeh, N. and Heller, K., 2021b, October. Deep Cox mixtures for survival regression. In *Machine Learning for Healthcare Conference* (pp. 674-708). PMLR.

**[8]** H. Ishwaran, Udaya B. Kogalur, Eugene H. Blackstone, and Michael S. Lauer. Random survival forests. The Annals of Applied Statistics, 2(3), 2008.

**[9]** Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. *arXiv preprint arXiv:1412.3555, 2014*.

**[10]** Sepp Hochreiter and J¨urgen Schmidhuber. Long short-term memory. *Neural computation*, 9 (8):1735–1780, 1997.

**[11]** LAG Ries, D Melbert, M Krapcho, DG Stinchcomb, N Howlader, MJ Horner, A Mariotto, BA Miller, EJ Feuer, SF Altekruse, et al. Seer cancer statistics review, 1975–2005. Bethesda, MD: National Cancer Institute, 2999, 2008.

**[12] **Nagpal, C., Goswami, M., Dufendach, K. and Dubrawski, A., 2022. Counterfactual Phenotyping with Censored Time-to-Events. *arXiv preprint arXiv:2202.11089*.

**[13]** W. A. Knaus, Harrell F. E., Lynn J, and et al. The support prognostic model: Objective estimates of survival for seriously ill hospitalized adults. Annals of Internal Medicine, 122: 191–203, 1995.

This article was initially published on the ML@CMU blog and appears here with the authors’ permission.

tags: deep dive

ML@CMU

Anna Demming delves into the details behind the prizes.

08 November 2024, by
Real World Data Science

Have you ever tried to use an AI tool like ChatGPT to do some math and found it doesn’t always add up?

07 November 2024, by
University of Waterloo

We introduce a new evaluation metric and benchmark dataset for automated evaluation of text-to-visual generative models.

06 November 2024, by
ML@CMU

We spoke to Amal about how AfriClimate AI started, and the projects and initiatives that the team are focussing on.

05 November 2024, by
Lucy Smith

A list of free-to-attend AI-related seminars that are scheduled to take place between 4 November and 31 December 2024.

04 November 2024, by
Lucy Smith

In this episode Ben chats to Lisa about data and AI literacy, data governance, ethical frameworks, and more.

01 November 2024, by
The Machine Ethics Podcast

©2021 - ROBOTS Association