Prediction of Survival Outcomes under Clinical Presence Shift: A Joint Neural Network Architecture
Vincent Jeanselme, Glen Martin, Matthew Sperrin, Niels Peek, Brian Tom, Jessica Barrett · 2025
Read Prediction of Survival Outcomes under Clinical Presence Shift: A Joint Neural Network Architecture by Vincent Jeanselme, Glen Martin, Matthew Sperrin, Niels Peek, Brian Tom, Jessica Barrett (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-05472.txt.
Introduction
Vincent Jeanselme and co-authors tackle health record data in a 2025 methods paper. Records reflect both illness and how often a patient meets the care system. That pattern of contact, called clinical presence, can shift when a model moves to a new site.
The team builds a neural network that learns visit timing and missing data alongside the survival outcome. Tests on a large hospital data set for death risk show gains in score and in travel to new settings. The work gives a formal name to presence shift. It also shows why models should learn the visit pattern, not ignore it.
Catalog record
| CREATOR | Vincent Jeanselme, Glen Martin, Matthew Sperrin, Niels Peek, Brian Tom, Jessica Barrett |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 74 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-05472
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Prediction of Survival Outcomes under Clinical Presence Shift: A Joint Neural Network Architecture Vincent Jeanselme, Glen Martin, Matthew Sperrin, Niels Peek, Brian Tom, Jessica Barrett Abstract Electronic health records arise from the complex interaction between patients and the healthcare system. This observation process of interactions, referred to as clinical presence , often impacts observed outcomes. When using electronic health records to develop clinical prediction models, it is standard practice to overlook clinical presence, impacting performance and limiting the transportability of models when this interaction evolves. We propose a multi-task recurrent neural network that jointly models the inter-observation time and the missingness processes characterising this interaction in parallel to the survival outcome of interest. Our work formalises the concept of clinical presence shift when the prediction model is deployed in new settings (e.g. different hospitals, regions or countries), and we theoretically justify why the proposed joint modelling can improve transportability under changes in clinical presence. We demonstrate, in a real-world mortality prediction task in the MIMIC-III dataset, how the proposed strategy improves performance and transportability compared to state-of-the-art prediction models that do not incorporate the observation process. These results emphasise the importance of leveraging clinical presence to improve performance and create more transportable clinical prediction models. ∎ 1 Introduction Due to their availability and cost-saving potential, Electronic Health Records (EHRs) are often considered a de facto resource for developing clinical prediction models goldstein2017opportunities . Unlike controlled experiments, EHRs reflect observational data arising from real-world interactions between patients and healthcare providers. These interactions are not random . Each data point in a health record reflects a conscious decision to collect information, based on the patient’s condition and the clinical staff’s assessment of what was needed and when. This observation process of interactions, which we refer to as clinical presence in the medical context, is informative as missingness and observation times are not independent of the patient’s conditions. When considering clinical presence in the prediction modelling context, current modelling strategies leverage some measures of this process to improve predictive performance sisk2020 . Indicators of missing tests, number of visits, and time since the last visit have been used as proxies for the patient’s condition, as temporality and missingness patterns often reflect the severity of a patient’s condition, but these proxies are ad hoc by nature and do not fully reflect the data-generating process. Instead of these approaches, practitioners can jointly model the clinical presence with the outcome process. Joint models sperrin2017informative ; gasparini2020mixed ; su2019sensitivity , marked point process islam2017marked and Markov models alaa2017learning model the visit process and the outcome of interest. Through shared effects, the data-generating process informs the modelling of the outcome of interest. However, these methods either rely on correct model specification with strong parametric assumptions or do not scale well to large datasets. Our work relies on neural networks to address these limitations. Critically, clinical presence processes differ across hospitals johnson2018generalizability , regions, and countries, as medical training and practice may differ and may be subject to policies subbaswamy2020development , insurance incentives, and the evolution of medical knowledge ghassemi2019practical . This clinical presence shift highlights a crucial limitation of clinical prediction models trained on EHRs; they may embed clinical presence patterns that do not transport under shift as practices and policies evolve van2020cautionary . Given the risk associated with the decisions these models inform, improving models’ transportability under these naturally occurring shifts is critical to applications in healthcare. We are unaware of work that has explored these models’ transportability under clinical presence shifts. Our work fills this gap by examining how handling irregular timings in longitudinal data may impact models’ transportability under clinical presence shifts. When considering potential shifts, modelling pipelines discard clinical presence patterns with the goal of improving transportability between observational settings. Although practitioners justify this practice to avoid leveraging a changeable and unpredictable process van2020cautionary ; xiao2018opportunities in the hope of improved transportability, evidence to back up or substantiate this claim in practice is lacking. On the contrary, we show that predictive performances decrease under shift when clinical presence is ignored. As a solution for improved transportability under clinical presence shifts, we propose a recurrent neural network that jointly models the time-to-event outcome of interest, inter-observation times and missingness features of a multivariate longitudinal process. Specifically, we model the time-to-event outcome using a neural network extension of the Cox model cox1972regression , known as DeepSurv katzman2018deepsurv , and, in parallel, the time to the subsequent observation and the missingness in laboratory tests of the multivariate longitudinal process. The proposed model, referred to as DeepJoint, enables scalable joint modelling by maximizing the full likelihood through end-to-end gradient descent. Our work is motivated by the prediction of mortality for patients admitted to intensive care, using data from the Medical Information Mart for Intensive Care III ( Mimic III ) johnson2016mimic . We use laboratory tests performed during the first 24 hours following admission as risk predictors for mortality after 24 hours. Of particular interest is the shift in the clinical presence patterns between patients admitted on weekends and weekdays. Through this natural experiment, our work explores how joint modelling improves the transportability of a model trained on patients admitted on weekdays to weekends. Our experiments show that the additional modelling of the time between successive orderings of laboratory tests and which out of a suite of 21 possible laboratory tests have been performed improves robustness to this shift in the observation process. Contrary to prevailing views, this experiment shows that not accounting for clinical presence not only reduces predictive performance but also harms transportability. By explicitly modelling clinical presence, DeepJoint results in a more transportable clinical prediction model in settings where the clinical presence process is likely to evolve. The outline of the paper is as follows. In Section 2 , we provide an introduction to the machine learning concepts used in the paper. In Section 3 , we review existing strategies to model clinical presence and improve transportability under shift. In Section 4 , we define the concept of clinical presence shift and consider transportability to different settings. In Section 5 , we propose the DeepJoint architecture and describe the training (estimation) of the model. In Section 6 , we present the results of the application to the Mimic III dataset, and we conclude with a discussion in Section 7 . 2 Background We first review the neural network components of our proposed architecture, which might be unfamiliar to statistical readers. After a brief introduction to neural networks, we describe the recurrent neural network structure used for handling longitudinal observations, the monotonic neural network, which we use to capture information about the time between successive observations (the inter-observation time), and the DeepSurv method, which we use for the survival outcome of interest. For an exhaustive description of neural networks and their evolution, please refer to bishop2006pattern ; goodfellow2016deep . 2.1 Neurons Neural networks aim to mimic biological brains composed of simple processing units — neurons — that, together, form complex systems. Formally, given a multidimensional input x\in\mathbb{R}^{m} with m\in\mathbb{N} its dimension, a neuron consists of a linear combination of its inputs — defined by a weight vector W\in\mathbb{R}^{m} and a bias 1 1 1 While this