Clinical clusters during acute illness in older frailer patients with long-term mortality predictive and inferential fidelity

Clinical clusters during acute illness in older frailer patients with long-term mortality predictive and inferential fidelity | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Clinical clusters during acute illness in older frailer patients with long-term mortality predictive and inferential fidelity Alex Tsui, Patrick Hogan, Hannah Cheston, Ophelia Dunne, Dominic Gardner, and 5 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6941703/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 29 Dec, 2025 Read the published version in BMC Medicine → Version 1 posted 9 You are reading this latest preprint version Abstract Background Defining acute illness decompensation as a single entity limits individualisation of treatments for older patients. Multi-modal and high-dimensional data offer opportunities to derive quantified clusters with clinically meaningful outcomes. We tested the hypothesis that cluster-driven and high-dimensional predictors can be constructed with sufficient fidelity for clinical deployment, concurrently highlighting mechanistic insights into pathophysiological substrates of acute illness decompensation, including where this affected the brain. Methods Two independent prospective cohort studies, DELPHIC and DECIDE, were harmonised and utilised as train and test partitions, contributing 209 and 205 unique first-participant acute admission episodes respectively. Baseline and acute illness variables were projected using T-stochastic neighbour embedding onto a two-dimensional manifold and agglomerative hierarchical clustering designated distance-defined subtypes. Predictive performances of clusters and full models were compared for brain decompensation within admission and two-year mortality. Shapley additive explanations (SHAPs) quantified directional contributions of inputs towards high-dimensional model performances. Results Three broad clinical subtypes were identified in older people during decompensation, with similar contributions from baseline and acute illness variables. From baseline to cluster-driven, and then high-dimensional prediction models for brain decompensation in admission, the test area under receiver operating characteristic curve (AUROC) improved from 0.563 to 0.641 and 0.797 respectively. Sleep-wake cycle disturbance was the most important predictor of delirium in admission, while physiological fluctuations within an admission episode, in particular from cognitive domains, were significant predictors of long-term mortality after acute admission. Conclusions Robust, generalisable clusters with clinical utility are discernable for older patients during acute illness. Our results demonstrate proof of concept for a longitudinal approach towards defining and modelling acute illness. We illustrate the potential to maximally predict adverse outcomes with high-dimensionality and multi-modality, and highlight the importance of sleep-wake cycle disturbances as a future target in studies of delirium neuropathohysiology. subtypes clusters prediction frailty geriatrics acute illness high-dimensional multi-modal modelling Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction When older patients are hospitalised with acute illness, a wide range of factors can impact clinical outcomes. This population is profoundly heterogeneous: there is large variation in how multimorbidity and functional impairments chronically accumulate, juxtapose with a myriad of acute aetiologies, and result in divergent patterns of decompensation to high-order functions such as sustained attention (delirium) or neuromuscular control of balance and gait (falls). In addition, presenting phenotypes are dynamic over the short and long-term, where there are differences in the pattern and severities of both the baseline (cognitive and physical function) and acute components (inattention, altered arousal, sleep-wake disturbance, other neuropsychiatric and motor deficits). Disaggregating these clinical variations requires detailed prospective studies capturing individuals before, during, and after acute illness [ 1 ], and studies mapping this in population samples have shown divergent outcomes [ 2 – 4 ]. Defining subtypes of acute decompensation is vital at both population and individual levels: robust and reproducible clusters with prognostic implications allow stratified approaches to calibrating management plans as well as future therapeutics development. At an individual patient level, while complex, high-dimensional models may produce the greatest predictive performances, particularly as data resolution and number of modalities increase, such algorithms have substantial computational demands which limits their scope for practical implementation. A dimensionality-reductive approach is an attractive compromise: algorithms using such techniques might yield marginally reduced yet comparable performance while offering sufficient scalability to be clinically and operationally effective. Previous models of acute illness in older patients have had poor or uncertain generalisability, most commonly due to a lack of external validation. With the advantage of data from prospective longitudinal population studies of older people in acute illness, contemporaneously recruited, the Delirium and Population Health Informatics Cohort (DELPHIC) [ 3 , 5 ] and Delirium and Cognitive Impact in Dementia (DECIDE) [ 2 ] studies offer the first externally validated data-driven longitudinal subtypes of acute illness in older people in inpatient settings. Second, we aimed to link identified subtypes with clinical outcomes. Third, we compared predictive performances of utilising cluster-driven and high-dimensional models, quantifying if the most accurate models demonstrated sufficient fidelity to deploy in clinical settings. Last, we considered mechanistic insights into pathological substrates for brain decompensation, highlighting potential targets for future studies. Methods Cohorts We used data from two independent British population cohorts, DELPHIC and DECIDE. Both have been described in detail [ 6 , 7 ]. DELPHIC was a longitudinal prospective sample, recruiting 1511 community-based participants aged ≥ 70 between 2018 and 2020 from the London Borough of Camden, with baseline assessments at recruitment, median follow-up at 3 years and daily assessments during any incident hospitalisation to University College London and Royal Free hospitals. DECIDE was a prospective sample nested in the Cognitive Function and Ageing Study II (CFAS-II), in which 1751 eligible participants ≥ 65 years were also assessed during each hospital admission to the Royal Victoria Infirmary or Freeman hospitals, Newcastle, between 5th January 2016 and 5th January 2017, followed by a repeat cognitive assessment 12 months after their most recent hospital admission. Data Harmonisation Variables in both cohorts were cross-mapped and harmonised. For the specific cognitive instruments, in DELPHIC, baseline cognition was based on TICS-m [ 8 ] and DECIDE used the mini-mental state examination (MMSE) and Cambridge Cognitive Examination (CAMCOG) [ 9 ]. Our primary outcomes were binary: cognitive decompensation (delirium within the admission episode) and two-year mortality. We harmonised a scaled score for orientation out of 3 based on itemised descriptions for each scale. For variables collected over multiple days of a hospital admission (e.g. MDAS items, heart rate, repeat blood tests), we summarised data across the admission episode using minimum, maximum, mean and standard deviation to offer a representative measure of cognition, function, physiological and biochemical pertubations. Prescribed medications were translated using World Health Organisation (WHO) Anatomical Therapeutic Chemical (ATC) classification system, defined as levels 1 to 5; we used level 2 (Supplementary Table 1). Missing data within each unique admission was filled forwards and backwards from any available data. Multiple imputation with chained equations was used for any continuous variables, missing drugs were filled as 0. Only data from the first chronological admission episode was used because within-person information from subsequent admissions would not be independent for test/train purposes. T-stochastic Neighbour Embedding (TSNE) and Hierarchical Clustering We constructed a matrix of Gower distances between each non-Bernoulli variable as well as calculating an alternative using standardisation and centering, with the assumption of all data being linear and continuous. Projecting these in two-dimensional spaces as a first step in visualising potential clusters, we used t-stochastic neighbour embedding (TSNE) [ 10 ], with interpolation points-based modular Python implementation (openTSNE) to define an approximate function and allowing addition of new data points onto an existing embedding [ 11 ], to compare how 2D projections from training and test datasets were distributed and that any potential clusters were not a consequence of data source alone. The first two-dimensional manifold was constructed using DELPHIC data only (perplexity: 20; metric: Euclidean); DECIDE data were added on to the same function and visually examined for similarity in space in a two-dimensional scatterplot (Supplementary Fig. 1). Clusters were delineated hierarchically using Euclidean distances between continuous variables with a “ward” definition of affinity, in which similarity was defined by the average squared distance between all pairs of datapoints. A horizontally drawn line across a dendrogram of hierarchical distances at the point of greatest vertical separation guided the optimal number of clusters, followed by manual inspection to ensure no cluster was composed of small numbers of data points. Interrogating feature contributions towards each cluster designation, predictors were univariately regressed against the cluster class. Similarly, for medication prescribed, the percentage of drugs prescribed for patients within each cluster was calculated Predictive modelling Survival up to 730 days (two years) was plotted in a Kaplan Meier graph by cohort and cluster designation. Difference in survival trajectories between clusters were analysed using log-rank test with cohort as an interaction term. We estimated predictive models of hierarchically increasing complexity, using first baseline variables of age and sex only, then cluster class and finally, full individual baseline and acute admission features. Targets were acute cognitive decompensation (delirium) and two-year mortality. XGBoost was selected for high-dimensional modelling due to its flexibility, ability to use mixed data, ease of hyperparameter optimisation and non-linear efficiency. In all models, data from DELPHIC were used for training and DECIDE for independent testing. Each partition contained only unique patients, with the first chronological admission episode selected for each patient. Ten-fold cross-validation was performed in the training DELPHIC dataset and optimised with area under the receiver operating characteristic curve (AUROC) as the evaluation metric. For the cognitive decompensation (delirium) model, AUROC was selected as the evaluation metric instead of precision recall receiver operating characteristic curve (PRAUC) despite an outcome imbalance, as it is clinically equally important to predict the absence of delirium as it is to predict the future presence of delirium. A calibration curve was also plotted to demonstrate predicted performance across a range of observed values. Manual grid searching determined the most optimal hyperparameters (number of estimators, maximum depth, minimum child weight, learning rate, gamma, subsample, column sample by tree (Supplementary Table 2), which were subsequently deployed on the independent test dataset. The weighted directional contribution of each feature to the final model was determined using SHapley Additive exPlanations (SHAP) values, with the most important twenty features visually demonstrated on a directional heatmap. Results Cohort demographics In DELPHIC, 209 participants were admitted over 372 number of admissions. 45% were diagnosed with delirium at any point during their included admission and 25% were no longer alive at 2 years after their first admission episode. In DECIDE, 205 number of participants were admitted over 318 number of admissions. 27% were diagnosed with delirium at any point and 22% were no longer alive at 2 years after their first admission episode. The mean ages for participants in DELPHIC and DECIDE were 80.6 and 81.6 years (p = 0.19) respectively. Baseline demographics were similar between both cohorts, although DELPHIC participants received more years of education (mean 11.3 compared with 10.1 years, p < 0.01). During acute illness, greater burden of delirium, arousal deficits and functional decline were demonstrated in DELPHIC patients, reflected in higher delirium incidence (DELPHIC 45%, DECIDE 26.8%, Table 1). However, acute illness severity as defined by NEWS2 was lower in DELPHIC (mean = 1.3) participants compared to DECIDE (mean = 5). Hierarchical clustering demonstrated three groups of 253, 123 and 38 participants in Clusters 1, 2 and 3 respectively. Distinct clusters could be spatially visualised in two-dimensional space (Fig. 1). DECIDE datapoints could be downsampled to a comparable 2-dimensional space using functionable representations from DELPHIC TSNE downsampling (Supplementary Fig. 1). Cluster interrogation Within each cluster, related groups of variables, such as baseline cognition measures and summarised measures of acute illness physiology, were correlated and trended in an anticipated manner. Individual clusters revealed unique clinical differences in baseline cognition, frailty, delirium burden, physiological derangements and premorbid prescriptions (Fig. 2a and Fig. 2b): Cluster 1 was composed of participants with higher measures of baseline cognition (MMSE item scores, fewer years of education), lower premorbid frailty (lower clinical frailty score, higher Barthel Index), lower burden of delirium during acute illness (lower MDAS and OSLA scores), biochemically demonstrated lower C-reactive protein and creatinine. Participants in Cluster 3 presented clinically orthogonally to Cluster 1 (lower baseline cognition, higher frailty, more delirium, higher CRP and lower albumin). Participants in Cluster 2 were clinically mid-range for baseline cognition, premorbid frailty, yet high in acute illness measures (delirium and arousal deficit scores, high CRP, high likelihood of physiological derangement). Participants in Cluster 1 received lower number of prescriptions of cardiovascular medications (antithrombotics, anti-anginals, beta blockers, diuretics, lipid lowering drugs), gastrointestinal medications (anatacids and laxatives), required less mineral supplementation, less analgesia, as well as less likely to receive psychoanaleptics. Participants in Cluster 3 demonstrated the highest prescription rates for all medications. There was no correlation between baseline cognition or frailty state with polypharmacy. Presence of analgesia and cardiovascular medications appeared most commonly in patients with polypharmacy. Cognitive decompensation prediction from clusters and longitudinally multimodal data XGB prediction models using only age and sex achieved a relatively poor AUROC of 0.563 on the independent test dataset. Addition of cluster designation achieved a significant increase in performance with AUROC of 0.641 while the best performing model, incorporating all individual overlapping baseline and acute illness admission features, achieved AUROC of 0.797 in an external test set (Fig. 3a). The best high-dimensional model demonstrated good calibration across the range of predicted and observed probabilities (Fig. 3b). SHAP values identified presence of pre-admission sleep-wake cycle disturbance (MDAS 10) as contributing most to predictive performance (Fig. 3c). Markers of underlying physiological, functional and cognitive function (hyponatraemia, low HABAM and total Barthel scores, lower number of years of education and poor baseline MMSE recall) plausibly related to underlying aetiologies, pre-existing prescriptions for constipation and pain relief were also predictive. Biochemical markers (CRP, platelets, haemoglobin and urea), all contributed information to prediction models but bidirectional in their importance, suggesting complex non-linear relationships with delirium incidence. Long-term mortality prediction from clusters and longitudinally multimodal data Differences in survival trajectories were demonstrated on Kaplan-Meier plots ((log-rank with cohort as interaction value, p value < 0.001, Fig. 4), with each analogous cluster in DELPHIC and DECIDE resulting in similar trajectories: participants in Cluster 1 have the most and Cluster 3 the least favourable survival trajectories respectively, with separation in trajectories appearing to initiate early from hospital discharge. Predictive modelling for two-year mortality demonstrated similar increasing performance with hierarchically more comp.ex inputs. Baseline age and sex only produced a AUROC of 0.572, improving to 0.664 with clusters, and 0.690 using full high dimensional inputs. SHAPS for the full model identified fluctuations of clinical, physiological and biochemical inputs within an admission (haemoglobin, potassium, MDAS 5: reduced ability to shift and maintain attention , heart rate standard deviation) to contribute the most to prediction. Poorer baseline cognitive fluency (number of animals named) and fewer years of education were poor prognostic factors. Pharmacologically, prescription of “drugs for acid related disorders” was a marker of increased mortality risk while proinflammatory state, as demonstrated by increased white cell count and CRP, increases two-year mortality risk. Discussion We applied unsupervised machine learning in two independent cohorts to demonstrate data-driven clusters within acute illness decompensation in older frailer patients. Clusters were prognostically associated with cognitive decompensation and two-year mortality. In addition, we quantified relative fidelities of reduced and high-dimensional approaches to predict adverse outcomes, objectively demonstrating potential performance of high-dimensional, multimodal and longitudinal prediction models. Taken together, our findings suggest different prognostic clusters are discernable, with implications for clinical management and inferring underlying disease mechanisms. Relative fidelities of deployable high-dimensional models and clinically informative low-dimensional clusters Consistent clusters across independent cohorts confirm complex yet coherent longitudinal subtypes, and that binary diagnoses such as “delirium” encompass multiple individual conditions, each of which requires detailed ascertainment to maximise individualised care. Clusters offer an objective framework to more optimally direct patient care towards specific therapies most likely to benefit, while in clinical trials, prognostic clusters augment participant stratification. Clusters were clinically informative beyond a description of age, sex, symptoms but the best predictive performances were produced by the full models, emphasising the added value of high-dimensionality. For delirium within admission at first contact, high-dimensional models achieved standards suitable for clinical deployment. Our model, derived from an unselected inpatient sample across non-intensive care settings, and externally validated, is unique. Other prediction models have been developed in post-operative or intensive care settings [ 12 , 13 ], but tend to lack validation in a test partition. Evolution of delirium definition and acute illness description Newly-described clinical clusters suggest a need to evolve current classification approaches: acute cognitive decompensation and long-term mortality prediction requires robust inputs ascertained from both pre-morbid baseline and acute illness states. However, this inherently longitudinal approach challenges contemporary delirium constructs: the DSM5 conception of delirium is essentially cross-sectional, with references to premorbid baseline defined simply as a ‘ change from baseline attention and awareness ’ [ 14 ]. Acute illness in older people is not well-captured by aggregated early warning scores [ 5 ]. Where physiological markers had large standard deviations, these were more prognostic than absolute measures. Moreover, the ranking of MDAS 5 ( reduced ability to shift and maintain attention ) suggest addition of specified cognitive domains, beyond “alert, voice, pain, unresponsive”, would offer easily implementable and feasible improvements on existing measures for highlighting acute illness severity. Examining the features with the most prognositic information, there are mechanistic inferences that cannot simply be extended from younger cohorts. For example, while systemically proinflammatory states such as sepsis are associated with mortality in younger patients, this is not the case in our sample. The impact of polypharmacy emerges with more nuance in our models. Survival trajectories between Cluster 1 and 3 diverged despite both groups receiving high prescription numbers, perhaps distinguishing subpopulations with appropriate versus inappropriate polypharmacy. Sleep-wake disturbance signifying a neurological substrate for delirium The contribution of MDAS 10 ( disturbed sleep-wake cycle ) towards our delirium prediction model is a well-recognised association. Yet, it is unclear whether our finding represents a subacute sleep-wake cycle disturbance suggestive of an early pre or subsyndromal presentation of delirium and/or an association with a longer-term circadian disruption indicating an inherent risk factor. The implication of circadian circuits being a neuropathophysiological substrate of delirium is consistent with studies on brainstem connectivity during delirium. Abnormal functional connectivity in mesencephalic, posteromedial cortex and brainstem ascending reticular activity system networks [ 15 ], and abnormal resting state connectivity in suprachiasmatic nuclei nodes and connections [ 16 ] have been demonstrated on fMRI investigations of medical inpatients experiencing incident delirium. In addition, the hypothalamic-pituitary-adrenal axis, which is intimately related in anatomy and function to circadian rhythms, is abnormally activated with blunted cortisol output in patients who have experienced delirium [ 17 , 18 ]. Isolation of relatively limited 10,000 suprachiasmatic nuclei neurones to be central neuropathophysiological substrates of delirium raises fundamental questions of what delirium is – does the phenotype simply reflect first decompensation of the most vulnerable part of complex brain networks? Is delirium a common final pathway of all cognitive dysfunctions, explaining why a spectrum of aetiologies can result in the same clinical syndrome, and when extended to known survival outcomes, therefore inherently simply a marker of cognitive frailty? Strengths and Limitations The principal advantage was the use of independent cohorts to test/train our models. Each cohort had comparable standards for outcome ascertainment and the models generalised well between samples despite some differences in baselines and acute illness presentations in the underlying cohorts Although we quantitatively derived clusters using distance-driven techniques, the process is dependent the number of modalities and dimensions employed. Pragmatically, the optimal number of subtypes is the number of distinct management options and prognoses, beyond which further separation have limited real-world impact. Second, our models utilised a limited number of modalities, namely clinical assessments and bloods, and a relatively small sample size for each cohort. A greater number of investigative modalities and deeper phenotyping, contributing more distinct pathophysiological information, would allow deployment of more advanced deep learning approaches, maximising predictive performance and hence inferential power, for delirium nor long-term mortality, with the aim of identifying candidate pathophysiological targets for future therapies. Conclusion In independent cohorts of older people with acute illness, we demonstrated structured subtypes using unsupervised machine learning. The clinical utility of these clusters, and the further performance gain when deploying a high-dimensional approach, are useful predictors of cognitive decompensation. Sleep-wake cycle disturbances may be central to cognitive decompensation. Our findings offer a proof-of-concept that clinical phenomenology in older frailer patients can be accurately characterised using data driven approaches. Declarations Ethics approval and consent to participate DELPHIC received approval from the Camden and King’s Cross Research Ethics Committee (16/LO/1217) and the Health Research Authority. DECIDE received approval from the North East - Newcastle & North Tyneside 2 Research Ethics Committee (15/NE/0353). Availability of data and materials Complete deidentified participant data, along with study protocols, consent forms, and case report forms are available through the Dementias Platform UK Data Portal: https://portal.dementiasplatform.uk/. Competing interests The authors declare that they have no competing interests Funding This study was funded by Wellcome Trust, Alzheimer’s Society UK, National Institute for Health Research, The Health Foundation, and EPSRC. The funders had no role in the design of the study; collection, analysis, and interpretation of data; and in writing the manuscript. Authors' contributions AT, PN and DD contributed to the conceptualisation of the study. AT, PH, HC, DD and SR curated the data. AT and PN did the formal analysis. AT, DD, SR acquired funding. AT, DD, PN, PH, HC, AM, DG and OD contributed to the methods. AT, DD, PN supervised the study. AT, DD and PN wrote the original draft. All authors reviewed and edited the manuscript. AT, DD and PN had access to, and could verify the data at all times. The corresponding author had full access to all the data and the final responsibility to submit for publication. Acknowledgements Not applicable Data sharing Complete deidentified participant data, along with study protocols, consent forms, and case report forms are available through the Dementias Platform UK Data Portal: https://portal.dementiasplatform.uk/. References MacLullich AMJ et al. Three key areas in progressing delirium practice and knowledge: recognition and relief of distress, new directions in delirium epidemiology and developing better research assessments. Age Ageing, 2022. 51(11). Richardson SJ, et al. Recurrent delirium over 12 months predicts dementia: results of the Delirium and Cognitive Impact in Dementia (DECIDE) study. Age Ageing. 2021;50(3):914–20. Tsui A, et al. The effect of baseline cognition and delirium on long-term cognitive impairment and mortality: a prospective population-based study. Lancet Healthy Longev. 2022;3(4):e232–41. Krogseth M, et al. Delirium, neurofilament light chain, and progressive cognitive impairment: analysis of a prospective Norwegian population-based cohort. Lancet Healthy Longev. 2023;4(8):e399–408. Tsui A et al. Extremes of baseline cognitive function determine the severity of delirium: a population study. Brain, 2023. Richardson SJ, et al. Protocol for the Delirium and Cognitive Impact in Dementia (DECIDE) study: A nested prospective longitudinal cohort study. BMC Geriatr. 2017;17(1):98. Davis D, et al. The delirium and population health informatics cohort study protocol: ascertaining the determinants and outcomes from delirium in a whole population. BMC Geriatr. 2018;18(1):45. Cook SE, Marsiske M, McCoy KJ. The use of the Modified Telephone Interview for Cognitive Status (TICS-M) in the detection of amnestic mild cognitive impairment. J Geriatr Psychiatry Neurol. 2009;22(2):103–9. de Koning I, et al. The CAMCOG: a useful screening instrument for dementia in stroke patients. Stroke. 1998;29(10):2080–6. van der Maaten L, Hinton G. Visualising data using t-SNE. J Mach Learn Res. 2008;9:2579–605. Poličar PGS, Zupan M. B.;, openTSNE: a modular Python library for t-SNE dimensionality reduction and embedding. J Stat Softw, 2024. 104(3). Lee DY, et al. Machine learning-based prediction model for postoperative delirium in non-cardiac surgery. BMC Psychiatry. 2023;23(1):317. Zhang Y, et al. Development of a machine learning-based prediction model for sepsis-associated delirium in the intensive care unit. Sci Rep. 2023;13(1):12697. American Psychiatric Association. The DSM-5 criteria, level of arousal and delirium diagnosis: inclusiveness is safer. BMC Med. 2014;12:141. Choi SH, et al. Neural network functional connectivity during and after an episode of delirium. Am J Psychiatry. 2012;169(5):498–507. Kyeong S, et al. Neural predisposing factors of postoperative delirium in elderly patients with femoral neck fracture. Sci Rep. 2018;8(1):7602. Cunningham C, Maclullich AM. At the extreme end of the psychoneuroimmunological spectrum: delirium as a maladaptive sickness behaviour response. Brain Behav Immun. 2013;28:1–13. Tsui A, et al. Longitudinal associations between diurnal cortisol variation and later-life cognitive impairment. Neurology. 2020;94(2):e133–41. Table 1 Table 1 is not available with this version. Additional Declarations No competing interests reported. Supplementary Files SupplementaryTable1.docx SupplementaryTable2.docx SupplementaryFig1.pdf Cite Share Download PDF Status: Published Journal Publication published 29 Dec, 2025 Read the published version in BMC Medicine → Version 1 posted Editorial decision: Revision requested 04 Aug, 2025 Reviews received at journal 24 Jul, 2025 Reviews received at journal 11 Jul, 2025 Reviewers agreed at journal 07 Jul, 2025 Reviewers agreed at journal 03 Jul, 2025 Reviewers invited by journal 02 Jul, 2025 Editor assigned by journal 23 Jun, 2025 Submission checks completed at journal 23 Jun, 2025 First submitted to journal 20 Jun, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6941703","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":480335309,"identity":"0dc05e78-aa7e-447d-9e92-987749adf6d5","order_by":0,"name":"Alex Tsui","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAt0lEQVRIiWNgGAWjYNACAxsGBnYokwjlzCB1aRCaBC0Mh0nQwt/Af/gzT8H5xP5mBsYPPxgOGxPUInGAmU2ax+B24ozDDMySPQyHzQi7C6iFGaRlA9Bh0kAX2hDUIX+Amfkzj8E5kBbm30RpMTjAzAB02AGQFjaQLYQdZniY2UxyjkGy8YzDjG2WPQbphL0vd7zx8Yc3f+xk+9ubD9/4UWFt2EBQDzOcxdhAXESOglEwCkbBKCAMAFetL99kBqBwAAAAAElFTkSuQmCC","orcid":"","institution":"UCL","correspondingAuthor":true,"prefix":"","firstName":"Alex","middleName":"","lastName":"Tsui","suffix":""},{"id":480335313,"identity":"56b8c1f9-b084-40aa-8ad5-9f3413fda069","order_by":1,"name":"Patrick Hogan","email":"","orcid":"","institution":"UCL","correspondingAuthor":false,"prefix":"","firstName":"Patrick","middleName":"","lastName":"Hogan","suffix":""},{"id":480335315,"identity":"8ba5da53-26ee-4aa6-9a9e-0d3ab32cffc2","order_by":2,"name":"Hannah Cheston","email":"","orcid":"","institution":"Central and North West London NHS Trust","correspondingAuthor":false,"prefix":"","firstName":"Hannah","middleName":"","lastName":"Cheston","suffix":""},{"id":480335318,"identity":"6468459c-3913-4fcd-8c89-16324be0db84","order_by":3,"name":"Ophelia Dunne","email":"","orcid":"","institution":"UCL","correspondingAuthor":false,"prefix":"","firstName":"Ophelia","middleName":"","lastName":"Dunne","suffix":""},{"id":480335319,"identity":"97716535-4827-4a6f-9def-a06f1f4ae55a","order_by":4,"name":"Dominic Gardner","email":"","orcid":"","institution":"UCL","correspondingAuthor":false,"prefix":"","firstName":"Dominic","middleName":"","lastName":"Gardner","suffix":""},{"id":480335320,"identity":"66404a29-1a11-4896-b244-eddead1410f6","order_by":5,"name":"Amy McWhirter","email":"","orcid":"","institution":"UCL","correspondingAuthor":false,"prefix":"","firstName":"Amy","middleName":"","lastName":"McWhirter","suffix":""},{"id":480335321,"identity":"14bc581e-0db2-4bb2-8b10-1e3015e291b2","order_by":6,"name":"Louise Allan","email":"","orcid":"","institution":"University of Exeter","correspondingAuthor":false,"prefix":"","firstName":"Louise","middleName":"","lastName":"Allan","suffix":""},{"id":480335322,"identity":"482b39ce-24ab-4357-9875-9af3de44f4c0","order_by":7,"name":"Sarah Richardson","email":"","orcid":"","institution":"Newcastle University","correspondingAuthor":false,"prefix":"","firstName":"Sarah","middleName":"","lastName":"Richardson","suffix":""},{"id":480335323,"identity":"594f466b-9763-49cd-bef6-b2023139d783","order_by":8,"name":"Parashkev Nachev","email":"","orcid":"","institution":"UCL","correspondingAuthor":false,"prefix":"","firstName":"Parashkev","middleName":"","lastName":"Nachev","suffix":""},{"id":480335325,"identity":"2636b309-c978-4b12-8024-d66b9e0dedde","order_by":9,"name":"Daniel Davis","email":"","orcid":"","institution":"UCL","correspondingAuthor":false,"prefix":"","firstName":"Daniel","middleName":"","lastName":"Davis","suffix":""}],"badges":[],"createdAt":"2025-06-20 22:23:12","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6941703/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6941703/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1186/s12916-025-04500-5","type":"published","date":"2025-12-29T15:57:15+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":86214779,"identity":"58dacccb-00d7-413c-bb81-3f7119b3c78b","added_by":"auto","created_at":"2025-07-08 05:52:02","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":32086,"visible":true,"origin":"","legend":"\u003cp\u003eLegend not included with this version.\u003c/p\u003e","description":"","filename":"1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6941703/v1/565481a7294a928316dd3025.jpg"},{"id":86215898,"identity":"f5eec6a2-772a-466a-b789-f84fafd72b9f","added_by":"auto","created_at":"2025-07-08 06:00:02","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":71352,"visible":true,"origin":"","legend":"\u003cp\u003eLegend not included with this version.\u003c/p\u003e","description":"","filename":"2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6941703/v1/824fb27a886b3f096579779f.jpg"},{"id":86214776,"identity":"2c2b8138-aca9-4d65-a263-c74daf22a3db","added_by":"auto","created_at":"2025-07-08 05:52:02","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":42847,"visible":true,"origin":"","legend":"\u003cp\u003eLegend not included with this version.\u003c/p\u003e","description":"","filename":"3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6941703/v1/0f2b743e64cabe2c955d1c24.jpg"},{"id":86216402,"identity":"1a9c915e-56a5-4721-ac90-588af75b9750","added_by":"auto","created_at":"2025-07-08 06:08:02","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":31611,"visible":true,"origin":"","legend":"\u003cp\u003eLegend not included with this version.\u003c/p\u003e","description":"","filename":"4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6941703/v1/96d92dfced558d4d06e40fd3.jpg"},{"id":99545317,"identity":"5219c8e1-3002-4159-b17f-395bdc3d0b63","added_by":"auto","created_at":"2026-01-05 16:05:50","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":940425,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6941703/v1/8448795a-94bd-4d4e-b876-c618437fca2d.pdf"},{"id":86215900,"identity":"bbfc2680-09dd-4e3d-83fd-a0bb81c03cc1","added_by":"auto","created_at":"2025-07-08 06:00:02","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":19766,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTable1.docx","url":"https://assets-eu.researchsquare.com/files/rs-6941703/v1/d4d49c7646e34415c821de36.docx"},{"id":86214778,"identity":"b999779f-5747-4f3d-a48a-1ff81ef6a18f","added_by":"auto","created_at":"2025-07-08 05:52:02","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":17717,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTable2.docx","url":"https://assets-eu.researchsquare.com/files/rs-6941703/v1/150117554a0cef4c352ecaa5.docx"},{"id":86214773,"identity":"0f8f0f19-9ac8-4af7-b71a-53dd1fc0a906","added_by":"auto","created_at":"2025-07-08 05:52:02","extension":"pdf","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":33991,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFig1.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6941703/v1/86b7f973c45ad0b9299f88f2.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Clinical clusters during acute illness in older frailer patients with long-term mortality predictive and inferential fidelity","fulltext":[{"header":"Introduction","content":"\u003cp\u003eWhen older patients are hospitalised with acute illness, a wide range of factors can impact clinical outcomes. This population is profoundly heterogeneous: there is large variation in how multimorbidity and functional impairments chronically accumulate, juxtapose with a myriad of acute aetiologies, and result in divergent patterns of decompensation to high-order functions such as sustained attention (delirium) or neuromuscular control of balance and gait (falls). In addition, presenting phenotypes are dynamic over the short and long-term, where there are differences in the pattern and severities of both the baseline (cognitive and physical function) and acute components (inattention, altered arousal, sleep-wake disturbance, other neuropsychiatric and motor deficits). Disaggregating these clinical variations requires detailed prospective studies capturing individuals before, during, and after acute illness [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e], and studies mapping this in population samples have shown divergent outcomes [\u003cspan additionalcitationids=\"CR3\" citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eDefining subtypes of acute decompensation is vital at both population and individual levels: robust and reproducible clusters with prognostic implications allow stratified approaches to calibrating management plans as well as future therapeutics development. At an individual patient level, while complex, high-dimensional models may produce the greatest predictive performances, particularly as data resolution and number of modalities increase, such algorithms have substantial computational demands which limits their scope for practical implementation. A dimensionality-reductive approach is an attractive compromise: algorithms using such techniques might yield marginally reduced yet comparable performance while offering sufficient scalability to be clinically and operationally effective.\u003c/p\u003e \u003cp\u003ePrevious models of acute illness in older patients have had poor or uncertain generalisability, most commonly due to a lack of external validation. With the advantage of data from prospective longitudinal population studies of older people in acute illness, contemporaneously recruited, the Delirium and Population Health Informatics Cohort (DELPHIC) [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e] and Delirium and Cognitive Impact in Dementia (DECIDE) [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e] studies offer the first externally validated data-driven longitudinal subtypes of acute illness in older people in inpatient settings. Second, we aimed to link identified subtypes with clinical outcomes. Third, we compared predictive performances of utilising cluster-driven and high-dimensional models, quantifying if the most accurate models demonstrated sufficient fidelity to deploy in clinical settings. Last, we considered mechanistic insights into pathological substrates for brain decompensation, highlighting potential targets for future studies.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eCohorts\u003c/h2\u003e \u003cp\u003eWe used data from two independent British population cohorts, DELPHIC and DECIDE. Both have been described in detail [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. DELPHIC was a longitudinal prospective sample, recruiting 1511 community-based participants aged\u0026thinsp;\u0026ge;\u0026thinsp;70 between 2018 and 2020 from the London Borough of Camden, with baseline assessments at recruitment, median follow-up at 3 years and daily assessments during any incident hospitalisation to University College London and Royal Free hospitals. DECIDE was a prospective sample nested in the Cognitive Function and Ageing Study II (CFAS-II), in which 1751 eligible participants\u0026thinsp;\u0026ge;\u0026thinsp;65 years were also assessed during each hospital admission to the Royal Victoria Infirmary or Freeman hospitals, Newcastle, between 5th January 2016 and 5th January 2017, followed by a repeat cognitive assessment 12 months after their most recent hospital admission.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eData Harmonisation\u003c/h3\u003e\n\u003cp\u003eVariables in both cohorts were cross-mapped and harmonised. For the specific cognitive instruments, in DELPHIC, baseline cognition was based on TICS-m [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e] and DECIDE used the mini-mental state examination (MMSE) and Cambridge Cognitive Examination (CAMCOG) [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Our primary outcomes were binary: cognitive decompensation (delirium within the admission episode) and two-year mortality. We harmonised a scaled score for \u003cem\u003eorientation\u003c/em\u003e out of 3 based on itemised descriptions for each scale. For variables collected over multiple days of a hospital admission (e.g. MDAS items, heart rate, repeat blood tests), we summarised data across the admission episode using minimum, maximum, mean and standard deviation to offer a representative measure of cognition, function, physiological and biochemical pertubations. Prescribed medications were translated using World Health Organisation (WHO) Anatomical Therapeutic Chemical (ATC) classification system, defined as levels 1 to 5; we used level 2 (Supplementary Table\u0026nbsp;1). Missing data within each unique admission was filled forwards and backwards from any available data. Multiple imputation with chained equations was used for any continuous variables, missing drugs were filled as 0. Only data from the first chronological admission episode was used because within-person information from subsequent admissions would not be independent for test/train purposes.\u003c/p\u003e\n\u003ch3\u003eT-stochastic Neighbour Embedding (TSNE) and Hierarchical Clustering\u003c/h3\u003e\n\u003cp\u003eWe constructed a matrix of Gower distances between each non-Bernoulli variable as well as calculating an alternative using standardisation and centering, with the assumption of all data being linear and continuous. Projecting these in two-dimensional spaces as a first step in visualising potential clusters, we used t-stochastic neighbour embedding (TSNE) [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], with interpolation points-based modular Python implementation (openTSNE) to define an approximate function and allowing addition of new data points onto an existing embedding [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e], to compare how 2D projections from training and test datasets were distributed and that any potential clusters were not a consequence of data source alone.\u003c/p\u003e \u003cp\u003eThe first two-dimensional manifold was constructed using DELPHIC data only (perplexity: 20; metric: Euclidean); DECIDE data were added on to the same function and visually examined for similarity in space in a two-dimensional scatterplot (Supplementary Fig.\u0026nbsp;1).\u003c/p\u003e \u003cp\u003eClusters were delineated hierarchically using Euclidean distances between continuous variables with a \u0026ldquo;ward\u0026rdquo; definition of affinity, in which similarity was defined by the average squared distance between all pairs of datapoints. A horizontally drawn line across a dendrogram of hierarchical distances at the point of greatest vertical separation guided the optimal number of clusters, followed by manual inspection to ensure no cluster was composed of small numbers of data points. Interrogating feature contributions towards each cluster designation, predictors were univariately regressed against the cluster class. Similarly, for medication prescribed, the percentage of drugs prescribed for patients within each cluster was calculated\u003c/p\u003e\n\u003ch3\u003ePredictive modelling\u003c/h3\u003e\n\u003cp\u003eSurvival up to 730 days (two years) was plotted in a Kaplan Meier graph by cohort and cluster designation. Difference in survival trajectories between clusters were analysed using log-rank test with \u003cem\u003ecohort\u003c/em\u003e as an interaction term.\u003c/p\u003e \u003cp\u003eWe estimated predictive models of hierarchically increasing complexity, using first baseline variables of age and sex only, then cluster class and finally, full individual baseline and acute admission features. Targets were acute cognitive decompensation (delirium) and two-year mortality. XGBoost was selected for high-dimensional modelling due to its flexibility, ability to use mixed data, ease of hyperparameter optimisation and non-linear efficiency. In all models, data from DELPHIC were used for training and DECIDE for independent testing. Each partition contained only unique patients, with the first chronological admission episode selected for each patient. Ten-fold cross-validation was performed in the training DELPHIC dataset and optimised with area under the receiver operating characteristic curve (AUROC) as the evaluation metric. For the cognitive decompensation (delirium) model, AUROC was selected as the evaluation metric instead of precision recall receiver operating characteristic curve (PRAUC) despite an outcome imbalance, as it is clinically equally important to predict the absence of delirium as it is to predict the future presence of delirium. A calibration curve was also plotted to demonstrate predicted performance across a range of observed values. Manual grid searching determined the most optimal hyperparameters (number of estimators, maximum depth, minimum child weight, learning rate, gamma, subsample, column sample by tree (Supplementary Table\u0026nbsp;2), which were subsequently deployed on the independent test dataset. The weighted directional contribution of each feature to the final model was determined using SHapley Additive exPlanations (SHAP) values, with the most important twenty features visually demonstrated on a directional heatmap.\u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eCohort demographics\u003c/h2\u003e \u003cp\u003eIn DELPHIC, 209 participants were admitted over 372 number of admissions. 45% were diagnosed with delirium at any point during their included admission and 25% were no longer alive at 2 years after their first admission episode. In DECIDE, 205 number of participants were admitted over 318 number of admissions. 27% were diagnosed with delirium at any point and 22% were no longer alive at 2 years after their first admission episode. The mean ages for participants in DELPHIC and DECIDE were 80.6 and 81.6 years (p\u0026thinsp;=\u0026thinsp;0.19) respectively. Baseline demographics were similar between both cohorts, although DELPHIC participants received more years of education (mean 11.3 compared with 10.1 years, p\u0026thinsp;\u0026lt;\u0026thinsp;0.01). During acute illness, greater burden of delirium, arousal deficits and functional decline were demonstrated in DELPHIC patients, reflected in higher delirium incidence (DELPHIC 45%, DECIDE 26.8%, Table\u0026nbsp;1). However, acute illness severity as defined by NEWS2 was lower in DELPHIC (mean\u0026thinsp;=\u0026thinsp;1.3) participants compared to DECIDE (mean\u0026thinsp;=\u0026thinsp;5).\u003c/p\u003e \u003cp\u003eHierarchical clustering demonstrated three groups of 253, 123 and 38 participants in Clusters 1, 2 and 3 respectively. Distinct clusters could be spatially visualised in two-dimensional space (Fig.\u0026nbsp;1). DECIDE datapoints could be downsampled to a comparable 2-dimensional space using functionable representations from DELPHIC TSNE downsampling (Supplementary Fig.\u0026nbsp;1).\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eCluster interrogation\u003c/h3\u003e\n\u003cp\u003eWithin each cluster, related groups of variables, such as baseline cognition measures and summarised measures of acute illness physiology, were correlated and trended in an anticipated manner. Individual clusters revealed unique clinical differences in baseline cognition, frailty, delirium burden, physiological derangements and premorbid prescriptions (Fig.\u0026nbsp;2a and Fig.\u0026nbsp;2b): Cluster 1 was composed of participants with higher measures of baseline cognition (MMSE item scores, fewer years of education), lower premorbid frailty (lower clinical frailty score, higher Barthel Index), lower burden of delirium during acute illness (lower MDAS and OSLA scores), biochemically demonstrated lower C-reactive protein and creatinine. Participants in Cluster 3 presented clinically orthogonally to Cluster 1 (lower baseline cognition, higher frailty, more delirium, higher CRP and lower albumin). Participants in Cluster 2 were clinically mid-range for baseline cognition, premorbid frailty, yet high in acute illness measures (delirium and arousal deficit scores, high CRP, high likelihood of physiological derangement). Participants in Cluster 1 received lower number of prescriptions of cardiovascular medications (antithrombotics, anti-anginals, beta blockers, diuretics, lipid lowering drugs), gastrointestinal medications (anatacids and laxatives), required less mineral supplementation, less analgesia, as well as less likely to receive psychoanaleptics. Participants in Cluster 3 demonstrated the highest prescription rates for all medications. There was no correlation between baseline cognition or frailty state with polypharmacy. Presence of analgesia and cardiovascular medications appeared most commonly in patients with polypharmacy.\u003c/p\u003e\n\u003ch3\u003eCognitive decompensation prediction from clusters and longitudinally multimodal data\u003c/h3\u003e\n\u003cp\u003eXGB prediction models using only age and sex achieved a relatively poor AUROC of 0.563 on the independent test dataset. Addition of cluster designation achieved a significant increase in performance with AUROC of 0.641 while the best performing model, incorporating all individual overlapping baseline and acute illness admission features, achieved AUROC of 0.797 in an external test set (Fig.\u0026nbsp;3a). The best high-dimensional model demonstrated good calibration across the range of predicted and observed probabilities (Fig.\u0026nbsp;3b). SHAP values identified presence of pre-admission sleep-wake cycle disturbance (MDAS 10) as contributing most to predictive performance (Fig.\u0026nbsp;3c). Markers of underlying physiological, functional and cognitive function (hyponatraemia, low HABAM and total Barthel scores, lower number of years of education and poor baseline MMSE recall) plausibly related to underlying aetiologies, pre-existing prescriptions for constipation and pain relief were also predictive. Biochemical markers (CRP, platelets, haemoglobin and urea), all contributed information to prediction models but bidirectional in their importance, suggesting complex non-linear relationships with delirium incidence.\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eLong-term mortality prediction from clusters and longitudinally multimodal data\u003c/h2\u003e \u003cp\u003eDifferences in survival trajectories were demonstrated on Kaplan-Meier plots ((log-rank with cohort as interaction value, p value\u0026thinsp;\u0026lt;\u0026thinsp;0.001, Fig.\u0026nbsp;4), with each analogous cluster in DELPHIC and DECIDE resulting in similar trajectories: participants in Cluster 1 have the most and Cluster 3 the least favourable survival trajectories respectively, with separation in trajectories appearing to initiate early from hospital discharge.\u003c/p\u003e \u003cp\u003ePredictive modelling for two-year mortality demonstrated similar increasing performance with hierarchically more comp.ex inputs. Baseline age and sex only produced a AUROC of 0.572, improving to 0.664 with clusters, and 0.690 using full high dimensional inputs. SHAPS for the full model identified fluctuations of clinical, physiological and biochemical inputs within an admission (haemoglobin, potassium, MDAS 5: \u003cem\u003ereduced ability to shift and maintain attention\u003c/em\u003e, heart rate standard deviation) to contribute the most to prediction. Poorer baseline cognitive fluency (number of animals named) and fewer years of education were poor prognostic factors. Pharmacologically, prescription of \u0026ldquo;drugs for acid related disorders\u0026rdquo; was a marker of increased mortality risk while proinflammatory state, as demonstrated by increased white cell count and CRP, increases two-year mortality risk.\u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eWe applied unsupervised machine learning in two independent cohorts to demonstrate data-driven clusters within acute illness decompensation in older frailer patients. Clusters were prognostically associated with cognitive decompensation and two-year mortality. In addition, we quantified relative fidelities of reduced and high-dimensional approaches to predict adverse outcomes, objectively demonstrating potential performance of high-dimensional, multimodal and longitudinal prediction models. Taken together, our findings suggest different prognostic clusters are discernable, with implications for clinical management and inferring underlying disease mechanisms.\u003c/p\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eRelative fidelities of deployable high-dimensional models and clinically informative low-dimensional clusters\u003c/h2\u003e \u003cp\u003eConsistent clusters across independent cohorts confirm complex yet coherent longitudinal subtypes, and that binary diagnoses such as \u0026ldquo;delirium\u0026rdquo; encompass multiple individual conditions, each of which requires detailed ascertainment to maximise individualised care. Clusters offer an objective framework to more optimally direct patient care towards specific therapies most likely to benefit, while in clinical trials, prognostic clusters augment participant stratification.\u003c/p\u003e \u003cp\u003eClusters were clinically informative beyond a description of age, sex, symptoms but the best predictive performances were produced by the full models, emphasising the added value of high-dimensionality. For delirium within admission at first contact, high-dimensional models achieved standards suitable for clinical deployment. Our model, derived from an unselected inpatient sample across non-intensive care settings, and externally validated, is unique. Other prediction models have been developed in post-operative or intensive care settings [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e], but tend to lack validation in a test partition.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eEvolution of delirium definition and acute illness description\u003c/h2\u003e \u003cp\u003eNewly-described clinical clusters suggest a need to evolve current classification approaches: acute cognitive decompensation and long-term mortality prediction requires robust inputs ascertained from both pre-morbid baseline and acute illness states. However, this inherently longitudinal approach challenges contemporary delirium constructs: the DSM5 conception of delirium is essentially cross-sectional, with references to premorbid baseline defined simply as a \u0026lsquo;\u003cem\u003echange from baseline attention and awareness\u003c/em\u003e\u0026rsquo; [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eAcute illness in older people is not well-captured by aggregated early warning scores [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Where physiological markers had large standard deviations, these were more prognostic than absolute measures. Moreover, the ranking of MDAS 5 (\u003cem\u003ereduced ability to shift and maintain attention\u003c/em\u003e) suggest addition of specified cognitive domains, beyond \u0026ldquo;alert, voice, pain, unresponsive\u0026rdquo;, would offer easily implementable and feasible improvements on existing measures for highlighting acute illness severity.\u003c/p\u003e \u003cp\u003eExamining the features with the most prognositic information, there are mechanistic inferences that cannot simply be extended from younger cohorts. For example, while systemically proinflammatory states such as sepsis are associated with mortality in younger patients, this is not the case in our sample. The impact of polypharmacy emerges with more nuance in our models. Survival trajectories between Cluster 1 and 3 diverged despite both groups receiving high prescription numbers, perhaps distinguishing subpopulations with appropriate versus inappropriate polypharmacy.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eSleep-wake disturbance signifying a neurological substrate for delirium\u003c/h2\u003e \u003cp\u003eThe contribution of MDAS 10 (\u003cem\u003edisturbed sleep-wake cycle\u003c/em\u003e) towards our delirium prediction model is a well-recognised association. Yet, it is unclear whether our finding represents a subacute sleep-wake cycle disturbance suggestive of an early pre or subsyndromal presentation of delirium and/or an association with a longer-term circadian disruption indicating an inherent risk factor. The implication of circadian circuits being a neuropathophysiological substrate of delirium is consistent with studies on brainstem connectivity during delirium. Abnormal functional connectivity in mesencephalic, posteromedial cortex and brainstem ascending reticular activity system networks [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e], and abnormal resting state connectivity in suprachiasmatic nuclei nodes and connections [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e] have been demonstrated on fMRI investigations of medical inpatients experiencing incident delirium. In addition, the hypothalamic-pituitary-adrenal axis, which is intimately related in anatomy and function to circadian rhythms, is abnormally activated with blunted cortisol output in patients who have experienced delirium [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. Isolation of relatively limited 10,000 suprachiasmatic nuclei neurones to be central neuropathophysiological substrates of delirium raises fundamental questions of what delirium is \u0026ndash; does the phenotype simply reflect first decompensation of the most vulnerable part of complex brain networks? Is delirium a common final pathway of all cognitive dysfunctions, explaining why a spectrum of aetiologies can result in the same clinical syndrome, and when extended to known survival outcomes, therefore inherently simply a marker of cognitive frailty?\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eStrengths and Limitations\u003c/h2\u003e \u003cp\u003eThe principal advantage was the use of independent cohorts to test/train our models. Each cohort had comparable standards for outcome ascertainment and the models generalised well between samples despite some differences in baselines and acute illness presentations in the underlying cohorts\u003c/p\u003e \u003cp\u003eAlthough we quantitatively derived clusters using distance-driven techniques, the process is dependent the number of modalities and dimensions employed. Pragmatically, the optimal number of subtypes is the number of distinct management options and prognoses, beyond which further separation have limited real-world impact. Second, our models utilised a limited number of modalities, namely clinical assessments and bloods, and a relatively small sample size for each cohort. A greater number of investigative modalities and deeper phenotyping, contributing more distinct pathophysiological information, would allow deployment of more advanced deep learning approaches, maximising predictive performance and hence inferential power, for delirium nor long-term mortality, with the aim of identifying candidate pathophysiological targets for future therapies.\u003c/p\u003e \u003c/div\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn independent cohorts of older people with acute illness, we demonstrated structured subtypes using unsupervised machine learning. The clinical utility of these clusters, and the further performance gain when deploying a high-dimensional approach, are useful predictors of cognitive decompensation. Sleep-wake cycle disturbances may be central to cognitive decompensation. Our findings offer a proof-of-concept that clinical phenomenology in older frailer patients can be accurately characterised using data driven approaches.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDELPHIC received approval from the Camden and King\u0026rsquo;s Cross Research Ethics Committee (16/LO/1217) and the Health Research Authority. DECIDE \u0026nbsp; received approval from the North East - Newcastle \u0026amp; North Tyneside 2 Research Ethics Committee (15/NE/0353).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eComplete deidentified participant data, along with study protocols, consent forms, and case report forms are available through the Dementias Platform UK Data Portal: https://portal.dementiasplatform.uk/.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was funded by Wellcome Trust, Alzheimer\u0026rsquo;s Society UK, National Institute for Health Research, The Health Foundation, and EPSRC. The funders had no role in the design of the study; collection, analysis, and interpretation of data; and in writing the manuscript.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026apos; contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAT, PN and DD contributed to the conceptualisation of the study. AT, PH, HC, DD and SR curated the data. AT and PN did the formal analysis. AT, DD, SR acquired funding. AT, DD, PN, PH, HC, AM, DG and OD contributed to the methods. AT, DD, PN supervised the study. AT, DD and PN wrote the original draft. All authors reviewed and edited the\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003emanuscript. AT, DD and PN had access to, and could verify the data at all times. The corresponding author had full access to all the data and the final responsibility to submit for publication.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData sharing\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eComplete deidentified participant data, along with study protocols, consent forms, and case report forms are available through the Dementias Platform UK Data Portal: https://portal.dementiasplatform.uk/.\u0026nbsp;\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eMacLullich AMJ et al. Three key areas in progressing delirium practice and knowledge: recognition and relief of distress, new directions in delirium epidemiology and developing better research assessments. Age Ageing, 2022. 51(11).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRichardson SJ, et al. Recurrent delirium over 12 months predicts dementia: results of the Delirium and Cognitive Impact in Dementia (DECIDE) study. Age Ageing. 2021;50(3):914\u0026ndash;20.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTsui A, et al. The effect of baseline cognition and delirium on long-term cognitive impairment and mortality: a prospective population-based study. Lancet Healthy Longev. 2022;3(4):e232\u0026ndash;41.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKrogseth M, et al. Delirium, neurofilament light chain, and progressive cognitive impairment: analysis of a prospective Norwegian population-based cohort. Lancet Healthy Longev. 2023;4(8):e399\u0026ndash;408.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTsui A et al. Extremes of baseline cognitive function determine the severity of delirium: a population study. Brain, 2023.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRichardson SJ, et al. Protocol for the Delirium and Cognitive Impact in Dementia (DECIDE) study: A nested prospective longitudinal cohort study. BMC Geriatr. 2017;17(1):98.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDavis D, et al. The delirium and population health informatics cohort study protocol: ascertaining the determinants and outcomes from delirium in a whole population. BMC Geriatr. 2018;18(1):45.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCook SE, Marsiske M, McCoy KJ. The use of the Modified Telephone Interview for Cognitive Status (TICS-M) in the detection of amnestic mild cognitive impairment. J Geriatr Psychiatry Neurol. 2009;22(2):103\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ede Koning I, et al. The CAMCOG: a useful screening instrument for dementia in stroke patients. Stroke. 1998;29(10):2080\u0026ndash;6.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003evan der Maaten L, Hinton G. Visualising data using t-SNE. J Mach Learn Res. 2008;9:2579\u0026ndash;605.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePoličar PGS, Zupan M. B.;, openTSNE: a modular Python library for t-SNE dimensionality reduction and embedding. J Stat Softw, 2024. 104(3).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee DY, et al. Machine learning-based prediction model for postoperative delirium in non-cardiac surgery. BMC Psychiatry. 2023;23(1):317.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang Y, et al. Development of a machine learning-based prediction model for sepsis-associated delirium in the intensive care unit. Sci Rep. 2023;13(1):12697.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAmerican Psychiatric Association. The DSM-5 criteria, level of arousal and delirium diagnosis: inclusiveness is safer. BMC Med. 2014;12:141.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChoi SH, et al. Neural network functional connectivity during and after an episode of delirium. Am J Psychiatry. 2012;169(5):498\u0026ndash;507.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKyeong S, et al. Neural predisposing factors of postoperative delirium in elderly patients with femoral neck fracture. Sci Rep. 2018;8(1):7602.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCunningham C, Maclullich AM. At the extreme end of the psychoneuroimmunological spectrum: delirium as a maladaptive sickness behaviour response. Brain Behav Immun. 2013;28:1\u0026ndash;13.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTsui A, et al. Longitudinal associations between diurnal cortisol variation and later-life cognitive impairment. Neurology. 2020;94(2):e133\u0026ndash;41.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"},{"header":"Table 1","content":"\u003cp\u003eTable 1 is not available with this version.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-medicine","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmed","sideBox":"Learn more about [BMC Medicine](http://bmcmedicine.biomedcentral.com/)","snPcode":"12916","submissionUrl":"https://submission.nature.com/new-submission/12916/3","title":"BMC Medicine","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"subtypes, clusters, prediction, frailty, geriatrics, acute illness, high-dimensional, multi-modal modelling","lastPublishedDoi":"10.21203/rs.3.rs-6941703/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6941703/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eDefining acute illness decompensation as a single entity limits individualisation of treatments for older patients. Multi-modal and high-dimensional data offer opportunities to derive quantified clusters with clinically meaningful outcomes. We tested the hypothesis that cluster-driven and high-dimensional predictors can be constructed with sufficient fidelity for clinical deployment, concurrently highlighting mechanistic insights into pathophysiological substrates of acute illness decompensation, including where this affected the brain.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eTwo independent prospective cohort studies, DELPHIC and DECIDE, were harmonised and utilised as train and test partitions, contributing 209 and 205 unique first-participant acute admission episodes respectively. Baseline and acute illness variables were projected using T-stochastic neighbour embedding onto a two-dimensional manifold and agglomerative hierarchical clustering designated distance-defined subtypes. Predictive performances of clusters and full models were compared for brain decompensation within admission and two-year mortality. Shapley additive explanations (SHAPs) quantified directional contributions of inputs towards high-dimensional model performances.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eThree broad clinical subtypes were identified in older people during decompensation, with similar contributions from baseline and acute illness variables. From baseline to cluster-driven, and then high-dimensional prediction models for brain decompensation in admission, the test area under receiver operating characteristic curve (AUROC) improved from 0.563 to 0.641 and 0.797 respectively. Sleep-wake cycle disturbance was the most important predictor of delirium in admission, while physiological fluctuations within an admission episode, in particular from cognitive domains, were significant predictors of long-term mortality after acute admission.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e \u003cp\u003eRobust, generalisable clusters with clinical utility are discernable for older patients during acute illness. Our results demonstrate proof of concept for a longitudinal approach towards defining and modelling acute illness. We illustrate the potential to maximally predict adverse outcomes with high-dimensionality and multi-modality, and highlight the importance of sleep-wake cycle disturbances as a future target in studies of delirium neuropathohysiology.\u003c/p\u003e","manuscriptTitle":"Clinical clusters during acute illness in older frailer patients with long-term mortality predictive and inferential fidelity","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-07-08 05:51:57","doi":"10.21203/rs.3.rs-6941703/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-08-04T08:30:54+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-07-25T01:24:03+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-07-11T22:48:45+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"47222992764067547158780945196424740923","date":"2025-07-07T22:37:29+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"200288741601193460105387407693167566554","date":"2025-07-03T09:21:58+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-07-02T14:35:56+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-06-23T05:47:44+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-06-23T05:36:39+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medicine","date":"2025-06-20T22:19:52+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medicine","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmed","sideBox":"Learn more about [BMC Medicine](http://bmcmedicine.biomedcentral.com/)","snPcode":"12916","submissionUrl":"https://submission.nature.com/new-submission/12916/3","title":"BMC Medicine","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"4abdea28-0207-4973-a6a8-5fb2977c1be5","owner":[],"postedDate":"July 8th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2026-01-05T16:01:02+00:00","versionOfRecord":{"articleIdentity":"rs-6941703","link":"https://doi.org/10.1186/s12916-025-04500-5","journal":{"identity":"bmc-medicine","isVorOnly":false,"title":"BMC Medicine"},"publishedOn":"2025-12-29 15:57:15","publishedOnDateReadable":"December 29th, 2025"},"versionCreatedAt":"2025-07-08 05:51:57","video":"","vorDoi":"10.1186/s12916-025-04500-5","vorDoiUrl":"https://doi.org/10.1186/s12916-025-04500-5","workflowStages":[]},"version":"v1","identity":"rs-6941703","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6941703","identity":"rs-6941703","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}