Statistical and Machine Learning Methods for Sequential Decision Making

This year’s Spring meeting will be organized in Rotterdam in honor of the 2022 Hans van Houwelingen Biometry Award won by Anirudh Tomer (Erasmus MC) for his paper with Dimitris Rizopoulos, Ewout Steyerberg, Daan Nieboer and Monique Roobol: “Shared decision making of burdensome surveillance tests using personalized schedules and their burden and benefit” published in 2022 in Statistics in Medicine.

Main location: Rotterdam, Erasmus MC

Location drinks: Erasmus MC 

Time: Thursday, 20 June 2024, 12:45h – 18:00h

Registration: registration is required using the form below.

Program: 

12.45 – 13.00: Coffee

13.00 – 13.45: BMS-ANed General Assembly (ALV)

13.45 – 14.00: Break

14.00 – 14.05: Opening

14.05 – 14.35: Anirudh Tomer (HvH award winner; Booking.com)

14.35 – 15.05: Zhenwei Yang (Erasmus Medical Center Rotterdam)

15.05 – 15.20: Break

15.20 – 16.10: Nina Deliu (Sapienza University of Rome)

16.10 – 17.00: Thomas Klausch (Amsterdam University Medical Centers)

17.00 – 18.00: Closing & Drinks

Abstracts:

Speaker: Anirudh Tomer (Booking.com)

Title: Personalized Schedules for Invasive Diagnostic Tests: A Panacea

Summary: Benchmark surveillance tests for detecting disease progression (e.g., biopsies, endoscopies) in early-stage chronic noncommunicable diseases (e.g., cancer, lung diseases) are usually burdensome. To detect progression on time, patients undergo invasive tests planned in a fixed, one-size-fits-all manner (e.g., annually). However, in recent years, a range of biometrics literature has been developed to tackle this problem. These methods vary in terms of the models they use (e.g., joint models for time-to-event and longitudinal data, Markov decision processes) and the criteria by which they evaluate their performance. Yet, applying these methods in the real world is still far from where it could be. In this talk, we aim to discuss these various methods and approaches and maintain a patient and doctor centric view to understand their applicability in the real world.

Speaker: Zhenwei Yang (Department of Biostatistics, Erasmus MC)

Title: A Bayesian Joint Modelling for Misclassified Interval-censoring and Competing Risks

Summary: Joint modeling of longitudinal and time-to-event outcomes (JM) has been applied in active surveillance in prostate cancer to study the association between longitudinal prostate-specific antigen (PSA) measurements and cancer progression. Prior work focused on the interval censoring of cancer progression due to periodic biopsies. However, there are two further challenges in this context: (1) patients initiating early treatment constitute a competing risk; (2) the biopsies can generate false-negative results, meaning that the underlying event could have happened earlier but was missed. To overcome misclassification and competing risks, we extended the Bayesian JM methodology to incorporate the biopsy sensitivity and a cause-specific survival submodel. Modeling the biopsy sensitivity is not straightforward since it is generally unknown and cannot be derived from the data at hand. Thus, the uncertainty of the biopsy sensitivity should also be considered. Under the Bayesian framework we follow, this uncertainty may be considered via a prior distribution, which may raise an identifiability issue. The issue can be resolved by specifying less flexible baseline hazards for the survival submodel.

Speaker: Nina Deliu (Department MEMOTEF Sapienza University of Rome)

Title: Online Sequential-Decision Making via Bandit Algorithms: Modeling Considerations for Better Decisions

Summary: The multi-armed bandit (MAB) framework holds great promise for optimizing sequential decisions online as new data arise. For example, it could be used to design adaptive experiments that can result in better participant outcomes and improved statistical power at the end of the study. However, due to mathematical and computational aspects, most MAB variants have been developed and are implemented under binary or normal outcome models. In this talk, guided by three biomedical case studies we have designed, I will illustrate how traditional statistics can be integrated within this framework to enhance its potential. Specifically, I will focus on the most popular Bayesian MAB algorithm, Thompson sampling, and two types of outcomes: (i) rating scales, increasingly common in recommendation systems, digital health, and education, and (ii) zero-inflated data, characterizing mobile health experiments. Theoretical properties and empirical advantages in terms of balancing exploitation (outcome performance) and exploration (learning performance) will be presented. Further considerations will be provided in the unique and challenging case of (iii) small samples. These works are the result of collaborations with Sofia Villar (Cambridge University), Bibhas Chakraborty (NUS University), and the IAI Lab (Toronto University), among others.

Speaker: Thomas Klausch (Department of Epidemiology and Data Science, Amsterdam University Medical Centers)

Title: BayesPIM: A Flexible Bayesian Prevalence-Incidence Mixture Model for Screening Data

Summary: We introduce a new model called BayesPIM (Bayesian prevalence-incidence mixture model), which is used to estimate disease incidence rates in populations that are screened for the occurrence of a disease, such as colorectal cancer or the pre-state of a disease, such as an adenoma. In the type of screening data that we consider, each subject has a series of interval-censored screening moments, where, at each moment, a test with imperfect sensitivity is administered. Due to the imperfect test, the result at each testing moment may be falsely negative with positive probability; for example, a colonoscopy can miss an adenoma. In addition, BayesPIM allows for the possibility that some individuals have the (pre-state) disease already at baseline, which is called prevalence. The prevalence status may be unobserved if the test at baseline is falsely negative or no baseline test is available. BayesPIM jointly models the disease incidence and baseline prevalence. Specifically, disease incidence is modeled by an accelerated failure time (AFT) model and disease prevalence is modeled by a probit model. In both models, covariates can be included to model heterogeneity over individuals and obtain stratified or personalized predictions of incidence and baseline prevalence. The key challenge in estimating the parameters of the joint model is dealing with the fact that the exact incidence time and prevalence status are unobserved due to interval censoring and imperfect testing. To tackle these challenges we develop a Gibbs algorithm that iterates over augmenting the latent incidence times and prevalence status, drawing the model parameters by Metropolis-Hastings, and drawing the sensitivity parameter from its full conditional distribution. We demonstrate the method’s efficacy in a simulation study and a case study of adenoma incidence in an international high-risk cohort screened for colorectal cancer. An implementation in R is available in the package BayesPIM.