Clinical Presence Shift in Healthcare ML

• Clinical presence shift is a type of distribution shift where the timing, frequency, and type of clinical measurements vary due to evolving patient interactions with the healthcare system.
• It involves modeling the longitudinal, inter-observation, and missingness processes through techniques like LSTM-based neural networks to predict survival outcomes more robustly.
• Incorporating clinical presence shift into models improves transportability across settings and helps address fairness concerns by capturing variations induced by different observational practices.

Searching arXiv for the cited papers and topic to ground the article in current research. Searching arXiv for "clinical presence shift DeepJoint (Jeanselme et al., 2022, Jeanselme et al., 7 Aug 2025)". Clinical presence shift is a form of distribution shift in clinical machine learning in which the observation process of interactions between patients and the healthcare system changes, thereby altering when measurements are made, which measurements are made, and, in some formulations, the values that become available for modeling. In this literature, clinical presence is not treated as a nuisance artifact but as an informative process: observational data in medicine arise from a complex interaction between patients and the healthcare system, and the sampling process is often highly irregular and itself informative. Models that ignore this process can show sub-optimal performance, limited generalisability, and reduced transportability when clinical practice, policy, or workflow evolves (Jeanselme et al., 2022, Jeanselme et al., 7 Aug 2025).

1. Definition and formal scope

Clinical presence refers to the interaction between patients and the healthcare system that determines which data are recorded in electronic health records. This includes patients’ health status and healthcare access, but also discretionary clinical decisions such as whether to order a test or record a variable. The resulting dataset therefore reflects not only biology but also institutional, social, and operational processes (Jeanselme et al., 2022).

A later formalization characterizes the observation process by the pair $(E, O)$, where $E$ is the time since the last encounter and $O$ is the indicator vector of covariates observed. Under this formulation, a clinical presence shift is a change in the distribution $q(E, O)$ while the concept distribution $q(Y \mid X, E, O)$ and the covariate distribution $q(X)$ are assumed unchanged. The consequence is that the observed covariate distribution $q(X^*)$ and the modeled $q(Y \mid X^*)$ shift even though the underlying outcome mechanism is held fixed (Jeanselme et al., 7 Aug 2025).

This definition distinguishes clinical presence shift from a generic train–test mismatch. The shift is specifically about how, when, and what gets measured. The motivating examples in the literature include changes across hospitals, policies, resource settings, and even weekday versus weekend admissions, where ordering frequency, timing, and missingness patterns differ despite comparable clinical tasks (Jeanselme et al., 2022).

2. Observation-process dimensions

In the DeepJoint line of work, clinical presence is decomposed into three time-varying and interrelated processes modeled alongside survival: the longitudinal process, the inter-observation process, and the missingness process. The longitudinal process concerns the actual laboratory test values at each observation. The inter-observation process concerns the timing between observations. The missingness process concerns which laboratory variables are measured or not measured at each time point (Jeanselme et al., 2022).

For the longitudinal process, an LSTM-derived hidden state $h_{i,j}$ parameterizes a neural network $L$ that predicts future laboratory values after a given interval. The probabilistic model is

$E$0

For the inter-observation process, a monotonic positive neural network $E$1 models the time until the next set of measurements through a cumulative-hazard parameterization,

$E$2

For the missingness process, a neural network $E$3 predicts the probability that each laboratory variable will be ordered after an interval $E$4,

$E$5

All three processes are predicted in parallel and contribute to training through a shared representation (Jeanselme et al., 2022).

A later formalization narrows the architectural emphasis to the observation-process variables most directly tied to the shift definition—inter-observation time and missingness—modeled jointly with the survival outcome. This suggests a distinction between a broad empirical treatment of clinical presence, which includes longitudinal values, and a narrower theoretical treatment centered on $E$6 (Jeanselme et al., 7 Aug 2025).

3. Joint survival modeling and transportability

DeepJoint uses a multi-task recurrent neural network in which an LSTM encoder processes irregular EHR sequences and produces latent states used by task-specific heads. The survival component is a DeepSurv-style non-linear proportional hazards model,

$E$7

where $E$8 is a neural network acting on the shared hidden state. In the 2022 formulation, training combines survival log-likelihood with likelihood terms for the longitudinal, temporal, and missingness processes; final fine-tuning of the survival head is then performed to maximize outcome discrimination (Jeanselme et al., 2022).

The 2025 formulation provides an explicit transportability argument for this joint modeling strategy. Its multi-task objective combines survival loss with losses for inter-observation time and missingness, using dynamic weighting during end-to-end training. The accompanying theoretical justification adapts multi-task robustness theory: when multiple related tasks are modeled jointly, the effect of small input shifts can be reduced, particularly when the task gradients are not fully redundant. In this account, modeling clinical presence regularizes the latent representation so that it captures both patient state and the observation regime, rather than exploiting correlations specific to one training environment (Jeanselme et al., 7 Aug 2025).

The central methodological claim is therefore not merely that missingness or timing should be appended as features, but that the observation process should be modeled as an output-generating process in its own right. This is the technical core of the argument that joint modeling can improve transportability under changes in clinical presence.

4. Empirical evidence from survival prediction

The principal empirical case study uses MIMIC-III laboratory data for mortality prediction after ICU admission. In the 2022 experiments, explicit modeling of longitudinal values, inter-observation times, and missingness improved performance relative to state-of-the-art predictive baselines, with DeepJointFineTune achieving a C-index at 1 day horizon of $E$9 (Jeanselme et al., 2022).

Robustness was assessed through a natural shift in clinical presence induced by weekday versus weekend admission. Models were trained on one setting and evaluated on the other while controlling for patient mix. The reported pattern was that models which only featurize timing or missingness, or ignore these patterns altogether, can perform well in-distribution yet lose discrimination under presence shift. By contrast, explicit joint modeling produced performance that was more stable under cross-setting transfer, and Brier scores were similar across methods, indicating no loss of calibration (Jeanselme et al., 2022).

The later mortality study reports transfer loss in C-index when moving from weekday-trained to weekend-tested models. DeepJoint recorded losses of $O$0 at 7 days, $O$1 at 30 days, and $O$2 for the integrated metric. The corresponding values were $O$3, $O$4, and $O$5 for Feature; $O$6, $O$7, and $O$8 for GRU-D; $O$9, $q(E, O)$0, and $q(E, O)$1 for Ignore; and $q(E, O)$2, $q(E, O)$3, and $q(E, O)$4 for Count. The same study states that each clinical presence dimension improves transportability to some extent, but joint modeling of both inter-observation time and missingness yields the most robust result (Jeanselme et al., 7 Aug 2025).

These findings are consistent with the broader claim that short-horizon deterioration prediction is especially sensitive to observation-process dynamics. Rapid changes in testing patterns are themselves informative for near-term risk, and models that encode this interaction explicitly are less vulnerable when the measurement regime changes.

5. Fairness, missingness, and group-specific mechanisms

Clinical presence shift has also been examined through the lens of algorithmic fairness. In that literature, clinical presence describes the way societal and decision biases shape what is absent from healthcare data, and clinical presence shift denotes changes or differences in the mechanisms of data observation across groups, settings, or time. The fairness analysis formalizes three clinical presence-driven missingness mechanisms: S1: Limited access to quality care, in which disadvantaged groups have less access and therefore more missing data; S2: (Mis)-informed collection, in which standard protocols fail to capture manifestations specific to marginalized groups; and S3: Confirmation bias, in which practitioners collect data based on expectations or stereotypes (Jeanselme et al., 2022).

The main theoretical result is that the fairness implications of imputation are underdetermined by observed data alone. The paper compares population mean imputation and group-specific mean imputation and shows that the expected reconstruction error depends on the underlying distribution and the unobservable missingness process. It further defines a fairness gap as the difference in reconstruction error between a group and its complement. Under this analysis, no universal imputation strategy guarantees smaller fairness or performance gaps across all clinical presence scenarios, and group-specific imputation can, under plausible mechanisms such as S3, increase the error or fairness gap for marginalized groups (Jeanselme et al., 2022).

The empirical findings reinforce this point. In simulations and in a MIMIC-III case study, overall predictive performance could be similar across imputation strategies while group-specific AUC and false negative rate gaps differed substantially, and sometimes even changed sign. For Black versus non-Black patients, population mean imputation reduced the false negative rate gap compared with group mean imputation, whereas stratification by insurance status could reverse the preference. The paper therefore recommends upstream exploration of missingness mechanisms, explicit justification of assumptions, sensitivity analysis across multiple imputation strategies, and reporting of both overall and group-level fairness metrics (Jeanselme et al., 2022).

This literature broadens the meaning of robustness under clinical presence shift. The question is not only whether models transport across sites or time, but also whether choices about handling missingness amplify or mitigate disparities that are already embedded in the observation process.

6. Relation to other shift literatures and to other meanings of “presence”

Clinical presence shift belongs to a wider family of deployment failures caused by distributional change, but it is not interchangeable with the other shift types studied in medical machine learning. In medical imaging, recent work distinguishes prevalence shift, covariate shift, and mixed shifts, and proposes an unsupervised identification framework that combines output-based tests with feature-based tests using self-supervised encoders; the practical point is that detecting shift is not sufficient if the root cause of the shift is unknown (Roschewitz et al., 2024). In health insurance claims, a large-scale retrospective scan across 1,010 tasks found temporal shift in 9.7% of tasks at the population level and in some discovered sub-population for 93.0% of tasks, emphasizing that clinically important non-stationarity is widespread and often concentrated in subgroups (Ji et al., 2023). In clinical question answering, the CLIFT benchmark showed marked performance degradation under natural distribution shifts, with reported F1 losses of approximately 44% to 77% depending on model and clinical subset (Pal, 2023).

A common misconception is that “clinical presence” has a single meaning across healthcare AI. In EHR prediction and fairness studies, it denotes the observation process that governs recorded data. In another body of work, however, presence refers to interpersonal or relational phenomena. The Presence Ontology defines presence as the purposeful practice of awareness, focus, and attention with the intent to understand and connect with patients (Maitra et al., 2021). A conversational-HRI position paper introduces bounded relational presence as attentiveness, continuity, and responsiveness without personhood, therapeutic authority, or human equivalence (Fried et al., 16 May 2026). This indicates a distinct terminological branch: those studies address relational interaction and care-adjacent design, whereas clinical presence shift in survival modeling concerns informative observation and its change over deployment environments.

Across these literatures, the shared lesson is that deployment validity depends on modeling or at least auditing the process that generates data, not only the target prediction task. In the specific sense used in EHR-based prediction, clinical presence shift names a clinically grounded source of non-stationarity: the healthcare system changes what the model gets to see.