Clinical Presence in Healthcare
- Clinical Presence is a multifaceted concept encompassing relational connection, documentation labels, and observational processes in clinical settings.
- It is applied in diverse areas such as palliative care communication, oral case presentation training, and AI-driven feature detection in biomedical data.
- Research demonstrates that modeling clinical presence can enhance clinical decision-making, improve data transportability, and address fairness in observational data.
Clinical presence is a heterogeneous term in contemporary biomedical and computational literature. It can denote the domain of interpersonal connection in medicine, described as a purposeful relational practice of awareness, focus, and attention intended to understand and connect with patients; a binary determination that a predefined clinical concept is present in documentation; or the informative observation process by which patient-healthcare interactions determine what becomes recorded in the electronic health record (Maitra et al., 2021, Marchal et al., 19 Jun 2026, Jeanselme et al., 2022). Related work extends the term to palliative-care communication metrics, oral case presentation training, and care-adjacent conversational systems, while adjacent technical literatures use “presence” for clinically meaningful feature detection, observation, embodiment, and clinician-induced occlusion (Wang et al., 2024, Ouyang et al., 27 Jan 2026, Fried et al., 16 May 2026).
1. Semantic range and major research traditions
The recent literature uses “clinical presence” across several distinct research programs rather than as a single unified construct. In some works it refers to relational qualities of encounters; in others it refers to presence/absence labels over text or images; and in still others it denotes the observation process that shapes EHR data before modeling begins (Maitra et al., 2021, Marchal et al., 19 Jun 2026, Jeanselme et al., 2022).
| Usage | Definition in the cited literature | Representative work |
|---|---|---|
| Interpersonal connection in medicine | purposeful, relational practice of awareness, focus, and attention intended to understand and connect with patients | Presence Ontology (Maitra et al., 2021) |
| Communication metric in care | being fully “with” the patient through listening, silence, and avoidance of interruption | CommSense (Wang et al., 2024) |
| Closed-ontology documentation detection | deciding whether a predefined ALS-related concept is present anywhere in a note section | ALS term extraction (Marchal et al., 19 Jun 2026) |
| Observation process in EHRs | inter-observation time and missingness shaped by patient-healthcare interaction | DeepJoint (Jeanselme et al., 2022) |
| Care-adjacent AI interaction quality | accompaniment, attentiveness, continuity, and responsiveness without personhood | bounded relational presence (Fried et al., 16 May 2026) |
| Professional communication skill | clarity, focus, confidence, and control of the clinical storyline in oral case presentation | CaseMaster (Ouyang et al., 27 Jan 2026) |
This suggests a family resemblance rather than a single ontology: presence is repeatedly linked to what is made clinically perceptible, interactionally legible, or computationally observable. The main differences concern the unit of analysis—encounter, note section, time series, conversation, or image—and whether presence is treated as a human quality, a label, or a sampling mechanism.
2. Presence as interpersonal connection in clinical encounters
In the ontology literature, clinical presence is the domain of interpersonal connection in medicine. The Presence Ontology models it not as a provider trait or single behavior, but as an emergent property of the encounter shaped by persons, characteristics, actions, objects, factors, qualities, and emotions. Its development combined a broad social-science literature survey, a systematic review of 21,835 articles with 77 retained for detailed conceptual analysis, relational ethnography of 5 pilot and 27 full clinical encounters, and interviews with 40 medical and nonmedical professionals; the resulting ontology was formalized in OWL in Protégé and aligned with resources such as SNOMED CT and the Emotion Ontology without importing them (Maitra et al., 2021).
The ontology’s structure makes several distinctions that are central to later work. Communication is subdivided into verbal, nonverbal, and paraverbal forms; tools and environment are modeled as active contributors rather than background context; and qualities such as empathy, compassion, trust, and warmth are treated as relational qualities that can be produced in interaction rather than as properties of one participant alone. Evaluation of the ontology led to the addition of Cognitive Models, including patient explanatory models and provider caregiving approaches, because earlier versions did not adequately represent how participants think about illness and care (Maitra et al., 2021).
In palliative-care sensing research, presence is operationalized more behaviorally. CommSense defines presence through silence, active listening, interruption, and speech balance, using smartwatch audio, speech-to-text transcripts, speaker identification, silence detection, overlap detection, and paraphrase identification. The system’s rule logic treats good presence as encouraging silence, engaging in conversation less than 50% of the time, and detecting paraphrasing as evidence of active listening; bad presence is associated with interruptions. In a pilot study with clinician participants in a simulated setting, presence achieved accuracy $0.862$, precision $0.723$, and recall $0.703$ for good scripts, and accuracy $0.797$, precision $0.835$, and recall $0.558$ for bad scripts (Wang et al., 2024).
A recurrent misconception in this literature is that presence is equivalent to quietness. The CommSense formulation is narrower and more interactional: silence can be good or bad, and presence depends on making space, confirming understanding, paraphrasing, and avoiding domination of the exchange (Wang et al., 2024). The ontology work goes further by locating these behaviors within a broader relational ecology of environment, tools, timing, framily, and cognitive models (Maitra et al., 2021).
3. Presence without personhood in conversational systems
Care-adjacent HRI introduces a different but related formulation. "Designing for Being-With: Presence Without Personhood" argues that conversational systems increasingly generate social presence through linguistic fluency, emotional mirroring, and continuity across interactions, but that in health- and care-adjacent settings these qualities risk relational overreach. The paper’s central distinction is explicit: “presence ≠ personhood.” Presence is framed as accompaniment, attentiveness, continuity, and responsiveness, while personhood, therapeutic authority, empathy as an inner state, mutual obligation, and human equivalence are rejected as design targets (Fried et al., 16 May 2026).
The proposed design stance is bounded relational presence, summarized by the phrase “being-with without becoming.” In this model, systems should remain interactionally present while avoiding personhood inference, authority drift, therapeutic overclaim, intimacy escalation, and the impression of unlimited availability. The paper’s formal contrast between “being there” and “being-with” opposes availability to attentiveness, implied mutuality to asymmetric relation, persona continuity to role clarity, escalating intimacy to bounded engagement, implied responsibility to explicit limits, and presence as performance to presence as composition (Fried et al., 16 May 2026).
The paper therefore distinguishes several pairs that are often conflated in public discourse. Helpful presence is separated from personhood; empathic effects are separated from claims of genuine empathy; authority is separated from role clarity; therapeutic framing is separated from care-adjacent utility; and presence is separated from persona continuity. The practical consequence is that timing, memory, silence, refusal, re-entry, exit, and handoff are treated as design materials. Withdrawal is not a failure state but a communicative act, and abrupt blocks or rate limits are criticized because they can feel like rupture or abandonment (Fried et al., 16 May 2026).
Its evaluation criteria are correspondingly non-benchmark-centric: relational coherence, honesty of limits, quality of withdrawal, non-escalation of intimacy, contextual appropriateness, and safety legibility. This is not a deployed clinical system. It is a designerly framework for care-adjacent interaction that deliberately resists the inference that a fluent system is empathic, competent, authoritative, or quasi-therapeutic (Fried et al., 16 May 2026).
4. Clinical presence as professional communication and educational attainment
Medical education literature uses clinical presence in yet another sense: the ability to present, synthesize, and communicate a case in a way that sounds clinically grounded. "CaseMaster" treats oral case presentation as a core form of physician-to-physician communication and explicitly links OCP training to clinical presence, defined here as clarity, focus, confidence, and control of the clinical storyline. The educational problem is that novices often over-include irrelevant details, omit key findings, or apply SOAP inflexibly, which weakens the presentation’s clinical authority (Ouyang et al., 27 Jan 2026).
CaseMaster operationalizes this through a two-stage web system built with Vue 3 and Python Flask, using OpenAI gpt-4o. In the preparation stage, learners review structured patient records and draft a SOAP-based presentation with an embedded assistant offering actions such as searching key knowledge points, reviewing medical literature, checking content logic, assessing reasonableness, providing definitions or examples, and giving presentation suggestions. In the reflection stage, learners compare their transcribed oral response to a reference answer and receive rubric-based JSON feedback over History, Important Information, Physical Examination, Labs, Assessment and Plan, and General and Style (Ouyang et al., 27 Jan 2026).
The controlled study involved 12 medical students in a within-subject comparison against a baseline workflow of Google search, markdown editing, and optional unrestricted ChatGPT. The clearest quantitative improvement was differential diagnosis clarity: CaseMaster , versus baseline , $0.862$0, with $0.862$1 and $0.862$2. System satisfaction also favored CaseMaster, $0.862$3 versus $0.862$4, with $0.862$5. Human quality ratings showed Cohen’s $0.862$6, while LLM scoring agreement with expert averages reached $0.862$7, with 3 minor errors in 112 manually reviewed LLM-generated scoring items (Ouyang et al., 27 Jan 2026).
The educational significance is that presence here is not bedside empathy or conversational companionship. It is a professional communicative competence: selecting salient data, making a case, sequencing findings, and presenting differential diagnosis logic in a manner recognizable as clinically assured. The paper nevertheless cautions against replacement of human instruction, notes a small sample size, a 15-minute artificial preparation window, and restriction to orthopedic cases, and recommends that LLM support remain aligned with human teaching rather than substitute for it (Ouyang et al., 27 Jan 2026).
5. Presence as detectable signal in clinical text
In clinical NLP, “presence” often has a sharply delimited meaning: whether a fixed clinical concept is documented anywhere in a unit of text. The ALS term-extraction study defines clinical presence detection as a closed-ontology multilabel classification problem over note sections. Given a section, the system decides which predefined ALS-related clinical terms are present anywhere in the text, without extracting exact values. Presence means that a concept is mentioned or inferable as documented in the section, and each ontology item receives a binary present/absent label (Marchal et al., 19 Jun 2026).
The study used a closed 85-label ontology spanning manual muscle test scores, medications, supplements, onset symptoms, forced vital capacity, ALSFRS-R total and subscore labels, symptom onset date, first diagnosis date, EMG findings, research participation, respiratory measures, numeric-heavy fields, and negation/absence statements. Data came from 200 note-section observations drawn from 23 discharge summaries across 3 ALS patients. Before prompting, the pipeline recursively extracted section text, excluded tabular content, normalized section titles, removed control characters outside printable ASCII, standardized whitespace, and reduced repeated line breaks while preserving paragraph boundaries. Twenty-six open-source small LLMs were prompted with a few-shot instructional template designed to return minified JSON, which was then parsed, JSON-repaired if malformed, normalized, and mapped to a binary multilabel matrix (Marchal et al., 19 Jun 2026).
Manual validation established that prompt-only SLMs did not outperform the best non-generative baseline. The regex baseline achieved micro-F1 $0.862$8, Hamming loss $0.862$9, precision $0.723$0, and recall $0.723$1. The best SLM by micro-F1, Qwen3-4B-Instruct-2507, achieved micro-F1 $0.723$2, precision $0.723$3, and recall $0.723$4. The TF-IDF label-similarity baseline achieved micro-F1 $0.723$5, precision $0.723$6, and recall $0.723$7. Hammer2.1-7B performed strongly for ALSFRS-R subscore detection, but the general conclusion was label-specific method choice and a hybrid workflow rather than replacement of rule-based extraction (Marchal et al., 19 Jun 2026).
A related phenotyping paper treats presence as the structured detectability of medical attributes in unstructured notes. Using the N2C2 2006 smoking dataset, it compared a ScispaCy-based pipeline that extracted the 250 most common biomedical terms, reduced them to 7 principal components with PCA, and classified notes with KNN, SVM, or MLP, against a ClinicalBERT + LSTM baseline. The alternative pipeline underperformed in micro-F1—KNN $0.723$8, MLP $0.723$9, SVM $0.703$0, versus ClinicalBERT + LSTM $0.703$1—but required about 4 hours rather than about 24 hours and was positioned as a supplement when runtime, interpretability, or local deployment mattered (Daniel, 2023).
A common misunderstanding is that presence detection in notes is a synonym for concept normalization or value extraction. The ALS study explicitly rejects that framing: presence detection is a first-stage extraction step, not open-ended concept discovery and not exact-value capture (Marchal et al., 19 Jun 2026).
6. Clinical presence as observation process, shift, and fairness problem
In longitudinal EHR modeling, clinical presence names the observation process itself. "DeepJoint" argues that observational medical data arise from complex patient-healthcare interactions, so the sampling process is informative rather than neutral. It formalizes clinical presence as an observation process $0.703$2, with the consequence that $0.703$3. The model decomposes clinical presence into three dimensions: the longitudinal process, the missingness process, and the inter-observation process, learned jointly with survival through an LSTM encoder and task-specific heads for future lab values, missingness, time to next observation, and DeepSurv-style risk (Jeanselme et al., 2022).
On a MIMIC-III mortality task using 17 laboratory tests from 30,834 adults, DeepJointFineTune achieved the best reported discrimination on the random split, with C-index $0.703$4 at 1 day, $0.703$5 at 7 days, and $0.703$6 at 14 days. More importantly, models that explicitly modeled observation dynamics were more robust to weekday-versus-weekend changes in clinical practice. DeepJointFeature showed transferred-versus-oracle differences of $0.703$7 at 1 day, $0.703$8 at 7 days, and $0.703$9 at 14 days, supporting the claim that modeling clinical presence improves transportability (Jeanselme et al., 2022).
The later joint-architecture paper formalizes clinical presence shift as a change in $0.797$0, the distribution of inter-observation time and missingness, under otherwise stable $0.797$1 and $0.797$2. Because such shifts alter both the observed covariate distribution $0.797$3 and the learned $0.797$4, models that ignore the observation process may fail when deployed in settings with different staffing, workflow, or practice regimes. In weekend-versus-weekday transfer experiments on 31,692 MIMIC-III patients and 21 laboratory tests, DeepJoint obtained the smallest overall transfer loss, reported as $0.797$5, and ablations indicated that both temporal and missingness modeling contributed to transportability (Jeanselme et al., 7 Aug 2025).
This literature treats missingness as structured signal rather than nuisance. The fairness paper on imputation under clinical presence calls missingness produced by patient access, clinician decision-making, and disease manifestation clinical missingness and shows that group-specific imputation can be misguided. Across simulations and MIMIC III, methods with similar overall predictive performance sometimes produced materially different subgroup AUC and false-negative-rate disparities, and group-specific imputation could worsen prediction disparities for the very group it was meant to help (Jeanselme et al., 2022).
A parallel deployment literature shows that label observation can itself be altered by clinician response to a model. In the label-selection paper, clinician-in-the-loop deployment of low-yield laboratory alerts changed which labels were observed; naive AUROC on the observed population could undershoot actual performance by up to 20%. The proposed remedy combined injected randomization with inverse probability weighting so that monitoring targeted the full deployment population rather than the selected labeled subset (Corbin et al., 2022).
Taken together, these results reject the idea that clinical presence is merely missingness noise. In this strand of work, it is a clinically meaningful and operationally unstable process that can improve internal prediction, distort evaluation, reduce transportability, and alter algorithmic fairness if not explicitly modeled (Jeanselme et al., 2022, Jeanselme et al., 2022).
7. Adjacent technical uses in imaging, sensing, rehabilitation, and VR
Several adjacent literatures use “presence” in clinically relevant but non-identical senses. In dermoscopy, the task is the presence of clinical dermoscopic features within skin lesions. The fully convolutional network paper reformulated superpixel classification as multi-label semantic segmentation for four feature classes—pigment network, negative network, milia-like cysts, and streaks—and reported first place in the 2017 ISIC-ISBI Part 2 challenge with average AUROC $0.797$6. The paper’s broader methodological point was that official AUROC-based rankings could be misleading, motivating a fuzzy Jaccard index that ignores empty masks (Kawahara et al., 2017).
In echocardiography, presence appears in the clinical claim that left-ventricular volume and ejection fraction are used to determine the presence and severity of heart-related conditions. The contour-sampling paper proposed CASUS, a contour-based aleatoric uncertainty framework that predicts contour location uncertainty, samples plausible endocardial contours, and propagates uncertainty to area, volume, FAC, and EF. Temporal consistency across end-diastole and end-systole was especially important for FAC and EF, and contour-based methods such as CASUS+t were stronger than pixel-wise aleatoric baselines for clinical metric uncertainty calibration (Judge et al., 18 Feb 2025).
In rehabilitation tracking, clinician presence is not relational or inferential but physical: therapists and standing frames create permanent and transient occlusions around spinal cord injury patients during stimulation-assisted training. The RGB-D tracking paper addressed this by representing each mesh vertex through geodesic distances to anchor points and resolving ambiguity with multi-hypothesis tracking. The resulting method was reported as robust to both surface deformations and transient occlusions, making low-cost motion tracking more realistic in clinician-assisted rehabilitation settings (Li et al., 2017).
The VR literature uses presence in the standard immersive sense of “being there,” which is conceptually distinct from clinical presence in medicine. In a visuo-haptic EEG study, higher immersion increased self-presence but not physical presence, while prediction-error disruptions elicited ERP and oscillatory effects without consistent moderation by presence scores. This usage belongs to the psychology of embodiment and immersion rather than to interpersonal care, documentation, or EHR observation processes, although it demonstrates how the term “presence” migrates across methodological domains (Gehrke et al., 27 Oct 2025).
A plausible implication of this broader technical record is that “clinical presence” functions as a cross-disciplinary contact term linking relation, observation, and detection. Its precise meaning depends on whether the research target is an encounter, a communicative act, a note section, an observation process, an image-derived feature, or a physical scene in which clinicians themselves alter what can be sensed.
