Data-driven Clusters Research Articles

FDG-PET patterns associate with survival in patients with prion disease.

Prion disease classically presents with rapidly progressive dementia, leading to death within months of diagnosis. Advances in diagnostic testing have improved recognition of patients with atypical presentations and protracted disease courses, raising key questions surrounding the relationship between patterns of neurodegeneration and survival. We assessed the contribution of fluorodeoxyglucose (FDG-PET) imaging for this purpose. FDG-PET were performed in 40 clinic patients with prion disease. FDG-PET images were projected onto latent factors generated in an external dataset to yield patient-specific eigenvalues. Eigenvalues were input into a clustering algorithm to generate data-driven clusters, which were compared by survival time. Median age at FDG-PET was 65.3 years (range 23-85). Median time from FDG-PET to death was 3.7 months (range 0.3-19.0). Four data-driven clusters were generated, termed "Neocortical" (n = 7), "Transitional" (n = 12), "Temporo-parietal" (n = 13), and "Deep nuclei" (n = 6). Deep nuclei and transitional clusters had a shorter survival time than the neocortical cluster. Subsequent analyses suggested that this difference was driven by greater hypometabolism of deep nuclei relative to neocortical areas. FDG-PET-patterns were not associated with demographic (age and sex) or clinical (CSF total-tau, 14-3-3) variables. Greater hypometabolism within deep nuclei relative to neocortical areas associated with more rapid decline in patients with prion disease and vice versa. FDG-PET informs large-scale network physiology and may inform the relationship between spreading pathology and survival in patients with prion disease. Future studies should consider whether FDG-PET may enrich multimodal prion disease prognostication models.

Multiple sclerosis subgroups: Data-driven clusters based on patient-reported outcomes and a large clinical sample.

While standard clinical assessments provide great value for people with multiple sclerosis (PwMS), they are limited in their ability to characterize patient perspectives and individual-level symptom heterogeneity. To identify PwMS subgroups based on patient-reported outcomes (PROs) of physical, cognitive, and emotional symptoms. We also sought to connect PRO-based subgroups with demographic variables, functional impairment, hypertension and smoking status, traditional qualitative multiple sclerosis (MS) symptom groupings, and neuroperformance measurements. Using a cross-sectional design, we applied latent profile analysis (LPA) to a large database of PROs; analytic sample N = 6619). We identified nine distinct MS subtypes based on PRO patterns. The subtypes were primarily categorized into low, moderate, and high mobility impairment clusters. Approximately 70% of participants were classified in a low mobility impairment group, 10% in a moderate mobility impairment group, and 20% in a high mobility impairment group. Within these subgroups, several unexpected patterns were observed, such as high mobility impairment clusters reporting low non-mobility impairment. The present study highlights an opportunity to advance precision medicine approaches in MS. Combining PROs with data-driven methodology allows for a cost-effective and personalized characterization of symptom presentations. that can inform clinical practice and future research designs.

Using machine learning to gain insights into chronic disease multimorbidity: trends and patterns in British Columbia, Canada

ObjectiveThe goal of this project is to explore novel ways to assess chronic disease multimorbidity (co-occurrence of two or more conditions) trends and patterns in the population of British Columbia (BC), Canada. ApproachThis study included linked data from the BC population (~5M individuals) from 2001/02 to 2019/20. We analysed 25 chronic conditions, including 7 primary cancer subtypes. We report multimorbidity (MM) incidence, prevalence, and most common disease combinations. Further we explore temporal MM disease patterns using directed network analyses and extracted data-driven disease clusters with an unsupervised machine learning algorithm. ResultsThe age-standardized incidence of MM stayed relatively stable over the study period for males and decreased for females, while prevalence increased to approximately 1 in 4 individuals in BC in 2019/20 (from 19% to 27% of females, from 15% to 22% of males). Disease networks and clusters varied significantly by sex and age group, this presentation will highlight select disease network and cluster findings and discuss their implications for chronic disease surveillance. ConclusionsThe prevalence of multimorbidity continues to rise in BC. Using advanced analytics to understand disease co-occurrence patterns provides new insights above and beyond traditional epidemiological metrics to support health system planning and prevention efforts. ImplicationsChronic disease surveillance and research have historically operated with a single disease focus, which is not patient-centered and does not adequately account for the reality of multimorbidity for many people. This project is laying the foundation for enhanced chronic disease surveillance and monitoring in BC.

Clinical characteristics and complication risks in data-driven clusters among Chinese community diabetes populations.

Novel diabetes phenotypes were proposed by the Europeans through cluster analysis, but Chinese community diabetes populations might exhibit different characteristics. This study aims to explore the clinical characteristics of novel diabetes subgroups under data-driven analysis in Chinese community diabetes populations. We used K-means cluster analysis in 6369 newly diagnosed diabetic patients from eight centers of the REACTION (Risk Evaluation of cAncers in Chinese diabeTic Individuals) study. The cluster analysis was performed based on age, body mass index, glycosylated hemoglobin, homeostatic modeled insulin resistance index, and homeostatic modeled pancreatic β-cell functionality index. The clinical features were evaluated with the analysis of variance (ANOVA) and chi-square test. Logistic regression analysis was done to compare chronic kidney disease and cardiovascular disease risks between subgroups. Overall, 2063 (32.39%), 658 (10.33%), 1769 (27.78%), and 1879 (29.50%) populations were assigned to severe obesity-related and insulin-resistant diabetes (SOIRD), severe insulin-deficient diabetes (SIDD), mild age-associated diabetes mellitus (MARD), and mild insulin-deficient diabetes (MIDD) subgroups, respectively. Individuals in the MIDD subgroup had a low risk burden equivalent to prediabetes, but with reduced insulin secretion. Individuals in the SOIRD subgroup were obese, had insulin resistance, and a high prevalence of fatty liver, tumors, family history of diabetes, and tumors. Individuals in the SIDD subgroup had severe insulin deficiency, the poorest glycemic control, and the highest prevalence of dyslipidemia and diabetic nephropathy. Individuals in MARD subgroup were the oldest, had moderate metabolic dysregulation and the highest risk of cardiovascular disease. The data-driven approach to differentiating the status of new-onset diabetes in the Chinese community was feasible. Patients in different clusters presented different characteristics and risks of complications.

Clusters of long COVID among patients hospitalized for COVID-19 in New York City

BackgroundRecent studies have demonstrated that individuals hospitalized due to COVID-19 can be affected by “long-COVID” symptoms for as long as one year after discharge.ObjectivesOur study objective is to identify data-driven clusters of patients using a novel, unsupervised machine learning technique.MethodsThe study uses data from 437 patients hospitalized in New York City between March 3rd and May 15th of 2020. The data used was abstracted from medical records and collected from a follow-up survey for up to one-year post-hospitalization. Hospitalization data included demographics, comorbidities, and in-hospital complications. The survey collected long-COVID symptoms, and information on general health, social isolation, and loneliness. To perform the analysis, we created a graph by projecting the data onto eight principal components (PCs) and running the K-nearest neighbors algorithm. We then used Louvain’s algorithm to partition this graph into non-overlapping clusters.ResultsThe cluster analysis produced four clusters with distinct health and social connectivity patterns. The first cluster (n = 141) consisted of patients with both long-COVID neurological symptoms (74%) and social isolation/loneliness. The second cluster (n = 137) consisted of healthy patients who were also more socially connected and not lonely. The third cluster (n = 96) contained patients with neurological symptoms who were socially connected but lonely, and the fourth cluster (n = 63) consisted entirely of patients who had traumatic COVID hospitalization, were intubated, suffered symptoms, but were socially connected and experienced recovery.ConclusionThe cluster analysis identified social isolation and loneliness as important features associated with long-COVID symptoms and recovery after hospitalization. It also confirms that social isolation and loneliness, though connected, are not necessarily the same. Physicians need to be aware of how social characteristics relate to long-COVID and patient’s ability to cope with the resulting symptoms.

Disease Progression of Data-Driven Subtypes of Parkinson's Disease: 5-Year Longitudinal Study from the Early Parkinson's Disease Longitudinal Singapore (PALS) Cohort.

The detailed trajectory of data-driven subtypes in Parkinson's disease (PD) within Asian cohorts remains undisclosed. To evaluate the motor, non-motor symptom (NMS) progression among the data-driven PD clusters. In this 5-year longitudinal study, NMS scale (NMSS), Hospital Anxiety Depression Scale (HADS), and Epworth sleepiness scale (ESS) were carried out annually to monitor NMS progression. H& Y staging scale, MDS-UPDRS part III motor score, and postural instability gait difficulty (PIGD) score were assessed annually to evaluate disease severity and motor progression. Five cognitive standardized scores were used to assess detailed cognitive progression. Linear mixed model was performed to assess the annual progression rates of the longitudinal outcomes. Two hundred and six early PD patients, consisting of 43 patients in cluster A, 98 patients in cluster B and 65 subjects in cluster C. Cluster A (severe subtype) had significantly faster progression slope in NMSS Domain 3 (mood/apathy) score (p = 0.01), NMSS Domain 4 (perceptual problems) score (p = 0.02), NMSS Domain 7 (urinary) score (p = 0.03), and ESS Total Score (p = 0.04) than the other two clusters. Cluster A also progressed significantly in PIGD score (p = 0.04). For cognitive outcomes, cluster A deteriorated significantly in visuospatial domain (p = 0.002), while cluster C (mild subtype) deteriorated significantly in executive domain (p = 0.04). The severe cluster had significantly faster progression, particularly in mood and perceptual NMS domains, visuospatial cognitive performances, and postural instability gait scores. Our findings will be helpful for clinicians to stratify and pre-emptively manage PD patients by developing intervention strategies to counter the progression of these domains.

Postpartum glucose intolerance after gestational diabetes mellitus: tailored prediction according to data-driven clusters and BMI-categories.

To account for the heterogeneity of gestational diabetes (GDM), this study investigated tailored predictors during pregnancy and at 6-8 weeks postpartum of glucose intolerance (GI) at 1-year postpartum. We identified predictors according to data-driven clusters, analogous to the newly proposed diabetes classification, and for clinical ease also based on BMI-categories. This is a secondary analysis of the MySweetheart trial. It included 179 women with GDM who underwent a 75g oral glucose tolerance test and HbA1c measurement at 1-year postpartum. Predictors were determined according to: a) cluster analysis based on age, BMI, HOMA-IR and HOMA-B; and b) BMI-categories (normal weight [NW], and overweight/obesity [OW/OB]). We identified two clusters during pregnancy and at 6-8 weeks postpartum (for both time points an "insulin-resistant", and an "insulin-deficient" cluster). The "insulin-resistant" cluster was associated with a 2.9-fold (CI: 1.46-5.87; pregnancy) and 3.5-fold (CI: 1.63-7.52; at 6-8 weeks postpartum) increased risk of GI at 1-year postpartum. During pregnancy, the most relevant predictors of GI were history of previous GDM and fasting glucose for the "insulin-deficient" and NW category and HOMA-IR for the "insulin-resistant" and OW/OB category (all p ≤0.035). In the postpartum, predictors were more heterogenous and included the insulin-sensitivity-adjusted-secretion index and 1-h glucose in the "insulin-deficient" and NW women. In women with GDM, we identified "insulin-resistant" and "insulin-deficient" clusters with distinct risks of future GI. Predictors varied according to clusters or BMI-categories emphasizing the need for tailored risk assessments.

Artificial intelligence classifies primary progressive aphasia from connected speech.

Neurodegenerative dementia syndromes, such as primary progressive aphasias (PPA), have traditionally been diagnosed based, in part, on verbal and non-verbal cognitive profiles. Debate continues about whether PPA is best divided into three variants and regarding the most distinctive linguistic features for classifying PPA variants. In this cross-sectional study, we initially harnessed the capabilities of artificial intelligence and natural language processing to perform unsupervised classification of short, connected speech samples from 78 pateints with PPA. We then used natural language processing to identify linguistic features that best dissociate the three PPA variants. Large language models discerned three distinct PPA clusters, with 88.5% agreement with independent clinical diagnoses. Patterns of cortical atrophy of three data-driven clusters corresponded to the localization in the clinical diagnostic criteria. In the subsequent supervised classification, 17 distinctive features emerged, including the observation that separating verbs into high- and low-frequency types significantly improved classification accuracy. Using these linguistic features derived from the analysis of short, connected speech samples, we developed a classifier that achieved 97.9% accuracy in classifying the four groups (three PPA variants and healthy controls). The data-driven section of this study showcases the ability of large language models to find natural partitioning in the speech of patients with PPA consistent with conventional variants. In addition, the work identifies a robust set of language features indicative of each PPA variant, emphasizing the significance of dividing verbs into high- and low-frequency categories. Beyond improving diagnostic accuracy, these findings enhance our understanding of the neurobiology of language processing.

71-OR: Rates of Remission and Metabolic Responses in Data-Driven High-Risk Prediabetes Clusters

71-OR: Rates of Remission and Metabolic Responses in Data-Driven High-Risk Prediabetes Clusters

1694-P: Subphenotype-Dependent Benefits of Bariatric Surgery in Data-Driven Clusters of Individuals at Risk for Type 2 Diabetes

1694-P: Subphenotype-Dependent Benefits of Bariatric Surgery in Data-Driven Clusters of Individuals at Risk for Type 2 Diabetes

Long-term seizure diary tracking habits in clinical studies: Evidence from the Human Epilepsy Project

ObjectiveTo characterize seizure tracking patterns of people with focal epilepsy using electronic seizure diary entries, and to assess for risk factors associated with poor tracking. MethodsWe analyzed electronic seizure diary data from 410 participants with newly diagnosed focal epilepsy in the Human Epilepsy Project 1 (HEP1). Each participant was expected to record data each day during the study, regardless of seizure occurrence. The primary outcome of this post-hoc analysis was whether each participant properly tracked a seizure diary entry each day during their study participation. Using finite mixture modeling, we grouped patient tracking trajectories into data-driven clusters. Once defined, we used multinomial modeling to test for independent risk factors of tracking group membership. ResultsUsing over up to three years of daily seizure diary data per subject, we found four distinct seizure tracking groups: consistent, frequent at study onset, occasional, and rare. Participants in the consistent tracking group tracked a median of 92% (interquartile range, IQR: 82%, 99%) of expected days, compared to 47% (IQR:34%, 60%) in the frequent at study onset group, 37% (IQR: 26%, 49%) in the occasional group, and 9% (IQR: 3%, 15%) in the rare group. In multivariable analysis, consistent trackers had lower rates of seizure days per tracked year during their study participation, compared to other groups. SignificanceFuture efforts need to focus on improving seizure diary tracking adherence to improve quality of outcome data, particularly in those with higher seizure burden. In addition, accounting for missing data when using seizure diary data as a primary outcome is important in research trials. If not properly accounted for, total seizure burden may be underestimated and biased, skewing results of clinical trials.

Secondary analysis of newly diagnosed type 2 diabetes subgroups and treatment responses in the MARCH cohort

ObjectiveTo incorporate new clusters in the MARCH (Metformin and AcaRbose in Chinese patients as the initial Hypoglycemic treatment) cohort of newly diagnosed type 2 diabetes (T2D) patients and compare the anti-glycemic effects of metformin and acarbose across different clusters. MethodsK-means cluster analysis was performed based on six clinical indicators. The diabetic clusters in the MARCH cohort were retrospectively associated with the response to metformin and acarbose. ResultsA total of 590 newly diagnosed T2D patients were classified by data-driven clusters into the MARD (mild obesity-related diabetes) (34.1 %), MOD (mild obesity-related diabetes) (34.1 %), SIDD (severe insulin-deficient diabetes) (20.3 %) and SIRD (severe insulin-resistant diabetes) (11.5 %) subgroups at baseline. At 24 and 48 weeks, 346 participants had finished the follow-up. After the adjustment of age, gender, weight, baseline HbA1c, baseline fasting glucose and 2-h postprandial blood glucose (2hPG), metformin mainly decreased the fasting glucose (0.07 ± 0.89 vs −0.26 ± 0.83, P = 0.043) in the MARD subgroup presented with OGTT (oral glucose tolerance test) results compared with acarbose group at 24 weeks. Acarbose led to a greater decrease in 2hPG in the MOD subgroup compared with metformin group (0.08 ± 0.86 vs −0.24 ± 0.92, P = 0.037) at 24 weeks. There was a also significant interaction between cluster and treatment efficacy in HbA1c (glycated hemoglobin) reduction in metformin and acarbose groups at 24 and 48 weeks (pinteraction<0.001). ConclusionsMetformin and acarbose affected different metabolic variables depending on the diabetes subtype.

Clinical Manifestations.

It is unclear whether primary progressive apraxia of speech (PPAOS) and progressive agrammatic aphasia (PAA) represent the two ends of a clinical continuum or exist as distinct syndromic entities with specific pathologic/prognostic correlates. We aimed to characterize the main clinical features along the spectrum of progressive non-fluent speech-language disorders (nf-SLD) and explore the existence of data-driven symptom clusters with etiological/prognostic value. We included 98 participants presenting with progressive motor speech impairment and/or agrammatism, of which 43 had an autopsy-confirmed neuropathological diagnosis. Speech pathologists rated motor speech features indicative of dysarthria and AOS. Quantitative measures of expressive/receptive agrammatism were obtained and compared with healthy controls. Baseline and longitudinal disease severity was assessed with the Clinical Dementia Rating sum-of-boxes (CDR-SB). We explored the clustering tendency of the data to form robust symptom clusters and used principal component analysis to extract data-driven clinical components. The longitudinal CDR-SB change was estimated with linear mixed-effects models. Ninety-one participants fitted previously reported clinical profiles (69 AOS+agrammatism, 18 PPAOS and 4 PAA), while 7 remained unclassifiable and were characterized by nonfluent speech and dysarthria in the absence of apraxia of speech or agrammatism. None of the baseline clinical features discriminated between FTLD subgroups. The Hopkins statistic ruled out the existence of syndromic clusters in the whole sample. Three data-driven components explained 71% of the variance ([i]severity-agrammatism, [ii]AOS, and [iii]dysarthria). The component characterized by prominent dysarthria was more specific to patients with Progressive Supranuclear Palsy while the component characterized by severity-agrammatism predicted a faster CDR-SB increase. Our results suggest that nf-SLD represents a clinical continuum. Splitting this continuum into different clinical phenotypes based on its major clinical features does not improve clinical-pathological correlations, stressing the need for new clinical and biological markers. The term nf-SLD could be used to group together these speech-language phenotypes strongly associated with FTLD.

The DNA methylation landscape of the human oxytocin receptor gene (OXTR): data-driven clusters and their relation to gene expression and childhood adversity

The oxytocin receptor gene (OXTR) is of interest when investigating the effects of early adversity on DNA methylation. However, there is heterogeneity regarding the selection of the most promising CpG sites to target for analyses. The goal of this study was to determine functionally relevant clusters of CpG sites within the OXTR CpG island in 113 mother-infant dyads, with 58 of the mothers reporting childhood maltreatment (CM). OXTR DNA methylation was analyzed in peripheral/umbilical blood mononuclear cells. Different complexity reduction approaches were used to reduce the 188 CpG sites into clusters of co-methylated sites. Furthermore, associations between OXTR DNA methylation (cluster- and site-specific level) and OXTR gene expression and CM were investigated in mothers. Results showed that, first, CpG sections differed strongly regarding their statistical utility for research of individual differences in DNA methylation. Second, cluster analyses and Partial Least Squares (PLS) suggested two clusters consisting of intron1/exon2 and the protein-coding region of exon3, respectively, as most strongly associated with outcome measures. Third, cross-validated PLS regression explained 7% of variance in CM, with low cross-validated variance explained for the prediction of gene expression. Fourth, substantial mother-child correspondence was observed in correlation patterns within the identified clusters, but only modest correspondence outside these clusters. This study makes an important contribution to the mapping of the DNA methylation landscape of the OXTR CpG island by highlighting clusters of CpG sites that show desirable statistical properties and predictive value. We provide a Companion Web Application to facilitate the choice of CpG sites.

Characterization of data-driven geriatric syndrome clusters in older people with HIV: a Mexican multicenter cross-sectional study

As living with HIV has been proposed as a condition that may accelerate aging, the main objective of this work was to estimate the prevalence of geriatric syndromes (GS) among older Mexicans with HIV dwelling in the community. Secondly, to evaluate whether the accumulation of GS could be associated with an adverse HIV-related clinical profile, independent of chronological age. Multicenter, cross-sectional study including 501 community-dwelling people aged ≥50 years with HIV. The overall prevalence of nine selected GS and their cumulative number were estimated. An Age-Independent Cumulative Geriatric Syndromes scale (AICGSs) was constructed, and correlations between the AICGSs and HIV-related parameters assessed. Finally, k-mean clustering analyses were performed to test the secondary objective. Median age 56 (IQR: 53-61) years, 81.6% of men. Polypharmacy (74.8%), sensorial deficit (71.2%), cognitive impairment (53.6%), physical disability (41.9%), pre-frailty (27.9%), and falls (29.7%), were the more prevalent GS. A significant negative correlation was found between the AICGSs and normalized values of CD4+ nadir cell counts (r=-0.126; 95%: CI:-0.223 to-0.026, p<0.05). Similarly, a significant inverse adjusted association between the CD4+ nadir cells and the AICGSs was observed on linear regression analysis (β-0.058; 95%: CI:-0.109 to-0.007, p=0.03). Cluster analysis identified three differentiated groups varying by age, metabolic comorbidities, AICGSs, and HIV-related parameters. An elevated prevalence of GS was observed in the studied population. Moreover, the accumulation of GS was associated with adverse HIV-related profiles, independent of age. Thus, early detection and management of GS are crucial to promote healthier aging trajectories in people with HIV. This work was funded in part by the National Center for the Prevention and Control of HIV/AIDS in Mexico (CENSIDA)-National Ministry of Health.

Definition of early age at onset in bipolar disorder according to distinctive neurodevelopmental pathways: insights from the FACE-BD study.

Converging evidence suggests that a subgroup of bipolar disorder (BD) with an early age at onset (AAO) may develop from aberrant neurodevelopment. However, the definition of early AAO remains unprecise. We thus tested which age cut-off for early AAO best corresponds to distinguishable neurodevelopmental pathways. We analyzed data from the FondaMental Advanced Center of Expertise-Bipolar Disorder cohort, a naturalistic sample of 4421 patients. First, a supervised learning framework was applied in binary classification experiments using neurodevelopmental history to predict early AAO, defined either with Gaussian mixture models (GMM) clustering or with each of the different cut-offs in the range 14 to 25 years. Second, an unsupervised learning approach was used to find clusters based on neurodevelopmental factors and to examine the overlap between such data-driven groups and definitions of early AAO used for supervised learning. A young cut-off, i.e. 14 up to 16 years, induced higher separability [mean nested cross-validation test AUROC = 0.7327 (± 0.0169) for ⩽16 years]. Predictive performance deteriorated increasing the cut-off or setting early AAO with GMM. Similarly, defining early AAO below 17 years was associated with a higher degree of overlap with data-driven clusters (Normalized Mutual Information = 0.41 for ⩽17 years) relatively to other definitions. Early AAO best captures distinctive neurodevelopmental patterns when defined as ⩽17 years. GMM-based definition of early AAO falls short of mapping to highly distinguishable neurodevelopmental pathways. These results should be used to improve patients' stratification in future studies of BD pathophysiology and biomarkers.

Vigilance associates with the low-dimensional structure of fMRI data

The human brain exhibits rich dynamics that reflect ongoing functional states. Patterns in fMRI data, detected in a data-driven manner, have uncovered recurring configurations that relate to individual and group differences in behavioral, cognitive, and clinical traits. However, resolving the neural and physiological processes that underlie such measurements is challenging, particularly without external measurements of brain state. A growing body of work points to underlying changes in vigilance as one driver of time-windowed fMRI connectivity states, calculated on the order of tens of seconds. Here we examine the degree to which the low-dimensional spatial structure of instantaneous fMRI activity is associated with vigilance levels, by testing whether vigilance-state detection can be carried out in an unsupervised manner based on individual BOLD time frames. To investigate this question, we first reduce the spatial dimensionality of fMRI data, and apply Gaussian Mixture Modeling to cluster the resulting low-dimensional data without any a priori vigilance information. Our analysis includes long-duration task and resting-state scans that are conducive to shifts in vigilance. We observe a close alignment between low-dimensional fMRI states (data-driven clusters) and measurements of vigilance derived from concurrent electroencephalography (EEG) and behavior. Whole-brain coactivation analysis revealed cortical anti-correlation patterns that resided primarily during higher behavioral- and EEG-defined levels of vigilance, while cortical activity was more often spatially uniform in states corresponding to lower vigilance. Overall, these findings indicate that vigilance states may be detected in the low-dimensional structure of fMRI data, even within individual time frames.

Diabetes duration and types of diabetes treatment in data-driven clusters of patients with diabetes

This study aimed to cluster patients with diabetes and explore the association between duration of diabetes and diabetes treatment choices in each cluster. A Two-Step cluster analysis was performed on 1332 Chinese patients with diabetes based on six parameters (glutamate decarboxylase antibodies, age at disease onset, body mass index, glycosylated hemoglobin, homeostatic model assessment 2 to estimate β-cell function and insulin resistance). Associations between the duration of diabetes and diabetes treatment choices in each cluster of patients were analyzed using Kaplan-Meier survival curves and logistic regression models. The following five replicable clusters were identified: severe autoimmune diabetes (SAID), severe insulin-deficient diabetes (SIDD), severe insulin-resistant diabetes (SIRD), mild obesity-related diabetes (MOD), and mild age-related diabetes (MARD). There were significant differences in blood pressure, blood lipids, and diabetes-related complications among the clusters (all P < 0.05). Early in the course of disease (≤5 years), compared with the other subgroups, the SIRD, MOD, and MARD populations were more likely to receive non-insulin hypoglycemic agents for glycemic control. Among the non-insulin hypoglycemic drug options, SIRD had higher rates of receiving metformin, alpha-glucosidase inhibitor (AGI), and glucagon-like peptide-1 drug; the MOD and MARD groups both received metformin, AGI and sodium-glucose cotransporter 2 inhibitor (SGLT-2i) drug ratio was higher. While the SAID and SIDD groups were more inclined to receive insulin therapy than the other subgroups, with SAID being more pronounced. With prolonged disease course (>5 years), only the MOD group was able to accept non-insulin hypoglycemic drugs to control the blood sugar levels, and most of them are still treated with metformin, AGI, and SGLT-2i drugs. While the other four groups required insulin therapy, with SIDD being the most pronounced. Clustering of patients with diabetes with a data-driven approach yields consistent results. Each diabetes cluster has significantly different disease characteristics and risk of diabetes complications. With the development of the disease course, each cluster receives different hypoglycemic treatments.

Characterization of data-driven clusters in diabetes-free adults and their utility for risk stratification of type 2 diabetes

BackgroundThe prevention of type 2 diabetes is challenging due to the variable effects of risk factors at an individual level. Data-driven methods could be useful to detect more homogeneous groups based on risk factor variability. The aim of this study was to derive characteristic phenotypes using cluster analysis of common risk factors and to assess their utility to stratify the risk of type 2 diabetes.MethodsData on 7317 diabetes-free adults from Sweden were used in the main analysis and on 2332 diabetes-free adults from Mexico for external validation. Clusters were based on sex, family history of diabetes, educational attainment, fasting blood glucose and insulin levels, estimated insulin resistance and β-cell function, systolic and diastolic blood pressure, and BMI. The risk of type 2 diabetes was assessed using Cox proportional hazards models. The predictive accuracy and long-term stability of the clusters were then compared to different definitions of prediabetes.ResultsSix risk phenotypes were identified independently in both cohorts: very low-risk (VLR), low-risk low β-cell function (LRLB), low-risk high β-cell function (LRHB), high-risk high blood pressure (HRHBP), high-risk β-cell failure (HRBF), and high-risk insulin-resistant (HRIR). Compared to the LRHB cluster, the VLR and LRLB clusters showed a lower risk, while the HRHBP, HRBF, and HRIR clusters showed a higher risk of developing type 2 diabetes. The high-risk clusters, as a group, had a better predictive accuracy than prediabetes and adequate stability after 20 years.ConclusionsPhenotypes derived using cluster analysis were useful in stratifying the risk of type 2 diabetes among diabetes-free adults in two independent cohorts. These results could be used to develop more precise public health interventions.

Clustering of trauma patients based on longitudinal data and the application of machine learning to predict recovery

Predicting recovery after trauma is important to provide patients a perspective on their estimated future health, to engage in shared decision making and target interventions to relevant patient groups. In the present study, several unsupervised techniques are employed to cluster patients based on longitudinal recovery profiles. Subsequently, these data-driven clusters were assessed on clinical validity by experts and used as targets in supervised machine learning models. We present a formalised analysis of the obtained clusters that incorporates evaluation of (i) statistical and machine learning metrics, (ii) clusters clinical validity with descriptive statistics and medical expertise. Clusters quality assessment revealed that clusters obtained through a Bayesian method (High Dimensional Supervised Classification and Clustering) and a Deep Gaussian Mixture model, in combination with oversampling and a Random Forest for supervised learning of the cluster assignments provided among the most clinically sensible partitioning of patients. Other methods that obtained higher classification accuracy suffered from cluster solutions with large majority classes or clinically less sensible classes. Models that used just physical or a mix of physical and psychological outcomes proved to be among the most sensible, suggesting that clustering on psychological outcomes alone yields recovery profiles that do not conform to known risk factors.