Survival Prediction: Algorithms, Challenges and Applications (SP-ACA)
A survival analysis model attempts to estimate the time until a specified event will happen in the future (or some related survival measure), for an individual. While the event of interest is often the time to death of a patient, it can alternatively be the time to relapse, or until the failure of a mechanical system, or until an employee leaves a company, etc. The key challenge in learning effective survival models is that this time-to-event is “censored’’ for some training instances — eg, we know only a lower bound on that time for some instances, or only an upper bound, etc. This means standard regression techniques cannot be used directly to compute survival measures from such training data. This has lead to a wide range of survival models, that each use the features of an instance (eg, a patient), available at the “start” time, to produce some survival measure, which might be a risk score, the probability of survival to a specific future time (eg, 1 year), or the survival probability over all future times, etc.
This symposium will focus on approaches for learning models that estimate survival measures from “survival datasets”, which include censored instances. It will include a mix of invited sessions, contributed talks, discussion groups and poster sessions. We seek contributions from leading experts and active researchers from diverse fields including machine learning, healthcare, medicine, finance and engineering. We anticipate this will foster interdisciplinary collaborations and will catalyze the development of the next generation of the survival prediction algorithms.
This symposium focuses on the following themes, covering four aspects of survival prediction tasks:
(1) Novel Algorithms;
(2) Evaluation Metrics;
(3) Foundational Issues;
Novel Algorithms: We solicit submissions with new conceptual frameworks for survival prediction models under various machine learning paradigms, including traditional algorithms (censored SVR, random forests, etc.), deep neural networks, and probabilistic graphical models. We will consider algorithms that compute atemporal risk scores, single-time probabilities or entire survival distributions for either a group of observations or for individual instances. We also encourage submissions that describe ways to learn survival models from multimodal data, for example, a patient’s clinical features, electronic health records, longitudinal health markers, gene expression data, and imaging data (MRI/CT/histopathology, etc.) for medical and clinical research problems, and from time-dependent covariates. We encourage submissions that include comprehensive software packages with relevant use-cases.
Evaluation metrics: Provide new strategies for comparing survival prediction models, deal with limitations of the data (eg, censored and missing data), address model calibration and discrimination issues, and discuss model comparison strategies such as bootstrapping, cross-validation and out-of-distribution evaluations.
Foundational Issues: Identifying causal effects of multiple covariates on survival estimation is important for selecting relevant predictive variables in various real-world applications. For example, many healthcare projects try to identify the influence of change of treatment on survival probability of an individual or a diseased population. Competing risks, the conditions in which two or more causes (e.g. cancer and COVID-19) compete to influence an individual’s time-to-event (e.g. death), present additional covariates that should be considered for causal inference. Thus, the algorithms for causal inference or counterfactual reasoning under competing risks for survival estimation problems will be considered. We also encourage submissions that discuss uncertainty quantification for point estimates of various survival measures.
Applications: Survival estimation applications of interest include, but are not limited to, the following:
Medicine and Healthcare (eg time to death, remission, or disease relapse/progression estimation, treatment impact assessment, treatment regimen selection)
Manufacturing and Engineering (eg machine failure, attrition rate estimation)
Finance and Economics (eg credit risk, inflation, employment or unemployment assessment, customer churn)
Law Enforcement (eg recidivism estimation)
For more information, please see the web page at https://spaca.weebly.com/