Using item response theory with health system data to identify latent groups of patients with multiple health conditions

A novel scale for assessing the risk of low birthweight: Birthweight questionnaire

BACKGROUND

The birthweight of a newborn is critical to their health, development, and well-being. Previous studies that used maternal characteristics to predict birthweight did not employ a harmonised scale to assess the risk of low birthweight (LBW).

OBJECTIVE

The goal of this study was to develop a new instrument that uses items on a uniform scale to assess the risk of an LBW in a pregnant woman.

METHODS

Item response theory was employed to evaluate a similar existing scale, and some weaknesses were identified.

RESULTS

Based on the observed weaknesses of the existing scale, a new uniform scale was developed, which is a 3-point Likert scale consisting of seven items.

CONCLUSION

The scale, termed birthweight questionnaire, is a valuable tool for collecting data that could assist in assessing the risk of an LBW at every stage of pregnancy.

Social determinants of multimorbidity patterns: A systematic review

Social determinants of multimorbidity are poorly understood in clinical practice. This review aims to characterize the different multimorbidity patterns described in the literature while identifying the social and behavioral determinants that may affect their emergence and subsequent evolution. We searched PubMed, Embase, Scopus, Web of Science, Ovid MEDLINE, CINAHL Complete, PsycINFO and Google Scholar. In total, 97 studies were chosen from the 48,044 identified. Cardiometabolic, musculoskeletal, mental, and respiratory patterns were the most prevalent. Cardiometabolic multimorbidity profiles were common among men with low socioeconomic status, while musculoskeletal, mental and complex patterns were found to be more prevalent among women. Alcohol consumption and smoking increased the risk of multimorbidity, especially in men. While the association of multimorbidity with lower socioeconomic status is evident, patterns of mild multimorbidity, mental and respiratory related to middle and high socioeconomic status are also observed. The findings of the present review point to the need for further studies addressing the impact of multimorbidity and its social determinants in population groups where this problem remains invisible (e.g., women, children, adolescents and young adults, ethnic groups, disabled population, older people living alone and/or with few social relations), as well as further work with more heterogeneous samples (i.e., not only focusing on older people) and using more robust methodologies for better classification and subsequent understanding of multimorbidity patterns. Besides, more studies focusing on the social determinants of multimorbidity and its inequalities are urgently needed in low- and middle-income countries, where this problem is currently understudied.

Psychometric Evaluation of GHQ-12 as a Screening Tool for Psychological Impairment of Healthcare Workers Facing COVID-19 Pandemic

BACKGROUND

The General Health Questionnaire (GHQ) is a widely used tool, both in clinical and research settings, due to its brevity and easy administration. Researchers often adopt a dichotomous measurement method, considering a total score above or below a certain threshold. This leads to an extreme simplification of the gathered data and therefore to the loss of clinical details. In a multi-step evaluation study aimed at assessing health care workers' mental health during the Covid-19 pandemic, GHQ-12 proved to be the most effective tool to detect psychological distress compared to other scales adopted. These results led to deepen the understanding of GHQ-12 properties through a statistical study by focusing on items' properties and characteristics.

METHODS

GHQ-12 responses were analyzed using Item Response Theory (IRT), a suitable method for scale assessment. Instead of considering the single overall score, in which each item accounts equally, it focuses on individual items' characteristics. Moreover, IRT models were applied combined with the latent class (LC) analysis, aiming to the determination of subgroups of individuals according to their level of psychological distress.

RESULTS

GHQ-12 was administered to 990 health-care workers and responses were scored using the binary method (0-0-1-1). We applied the two-parameter logistic (2-PL) model, finding that the items showed different ways of responses and features. The latent class analysis classified subjects into three sub-groups according to their responses to GHQ-12 only: 47% of individuals with general well-being, 38% expressing signs of discomfort without severity and 15% of subjects with a high level of impairment. This result almost reproduces subjects' classification obtained after administering the six questionnaires of the study protocol.

CONCLUSIONS

Accurate statistical techniques and a deep understanding of the latent factors underlying the GHQ-12 resulted in a more effective usage of such psychometric questionnaire - i.e. a more refined gathering of data and a significant time and resource efficiency. We underlined the need to maximize the extraction of data from questionnaires and the necessity of them being less lengthy and repetitive.

Machine learning approaches for electronic health records phenotyping: a methodical review

OBJECTIVE

Accurate and rapid phenotyping is a prerequisite to leveraging electronic health records for biomedical research. While early phenotyping relied on rule-based algorithms curated by experts, machine learning (ML) approaches have emerged as an alternative to improve scalability across phenotypes and healthcare settings. This study evaluates ML-based phenotyping with respect to (1) the data sources used, (2) the phenotypes considered, (3) the methods applied, and (4) the reporting and evaluation methods used.

MATERIALS AND METHODS

We searched PubMed and Web of Science for articles published between 2018 and 2022. After screening 850 articles, we recorded 37 variables on 100 studies.

RESULTS

Most studies utilized data from a single institution and included information in clinical notes. Although chronic conditions were most commonly considered, ML also enabled the characterization of nuanced phenotypes such as social determinants of health. Supervised deep learning was the most popular ML paradigm, while semi-supervised and weakly supervised learning were applied to expedite algorithm development and unsupervised learning to facilitate phenotype discovery. ML approaches did not uniformly outperform rule-based algorithms, but deep learning offered a marginal improvement over traditional ML for many conditions.

DISCUSSION

Despite the progress in ML-based phenotyping, most articles focused on binary phenotypes and few articles evaluated external validity or used multi-institution data. Study settings were infrequently reported and analytic code was rarely released.

CONCLUSION

Continued research in ML-based phenotyping is warranted, with emphasis on characterizing nuanced phenotypes, establishing reporting and evaluation standards, and developing methods to accommodate misclassified phenotypes due to algorithm errors in downstream applications.

Multimorbidity Patterns and Their Association with Social Determinants, Mental and Physical Health during the COVID-19 Pandemic

BACKGROUND

The challenge posed by multimorbidity makes it necessary to look at new forms of prevention, a fact that has become heightened in the context of the pandemic. We designed a questionnaire to detect multimorbidity patterns in people over 50 and to associate these patterns with mental and physical health, COVID-19, and possible social inequalities.

METHODS

This was an observational study conducted through a telephone interview. The sample size was 1592 individuals with multimorbidity. We use Latent Class Analysis to detect patterns and SF-12 scale to measure mental and physical quality-of-life health. We introduced the two dimensions of health and other social determinants in a multinomial regression model.

RESULTS

We obtained a model with five patterns (entropy = 0.727): 'Relative Healthy', 'Cardiometabolic', 'Musculoskeletal', 'Musculoskeletal and Mental', and 'Complex Multimorbidity'. We found some differences in mental and physical health among patterns and COVID-19 diagnoses, and some social determinants were significant in the multinomial regression.

CONCLUSIONS

We identified that prevention requires the location of certain inequalities associated with the multimorbidity patterns and how physical and mental health have been affected not only by the patterns but also by COVID-19. These findings may be critical in future interventions by health services and governments.

Characteristics of High-Need, High-Cost Patients : A "Best-Fit" Framework Synthesis

BACKGROUND

Accurately identifying high-need, high-cost (HNHC) patients to reduce their preventable or modifiable health care use for their chronic conditions is a priority and a challenge for U.S. policymakers, health care delivery systems, and payers.

PURPOSE

To identify characteristics and criteria to distinguish HNHC patients.

DATA SOURCES

Searches of multiple databases and gray literature from 1 January 2000 to 22 January 2022.

STUDY SELECTION

English-language studies of characteristics and criteria to identify HNHC adult patients, defined as those with high use (emergency department, inpatient, or total services) or high cost.

DATA EXTRACTION

Independent, dual-review extraction and quality assessment.

DATA SYNTHESIS

The review included 64 studies comprising multivariate exposure studies (n = 47), cluster analyses (n = 11), and qualitative studies (n = 6). A National Academy of Medicine (NAM) taxonomy was an initial "best-fit" framework for organizing the synthesis of the findings. Patient characteristics associated with being HNHC included number and severity of comorbid conditions and having chronic clinical conditions, particularly heart disease, chronic kidney disease, chronic lung disease, diabetes, cancer, and hypertension. Patients' risk for being HNHC was often amplified by behavioral health conditions and social risk factors. The reviewers revised the NAM taxonomy to create a final framework, adding chronic pain and prior patterns of high health care use as characteristics associated with an increased risk for being HNHC.

LIMITATION

Little evidence distinguished potentially preventable or modifiable health care use from overall use.

CONCLUSION

A combination of characteristics can be useful for identifying HNHC patients. Because of the complexity of their conditions and circumstances, improving their quality of care will likely also require an individualized assessment of care needs and availability of support services.

PRIMARY FUNDING SOURCE

Agency for Healthcare Research and Quality. (PROSPERO: CRD42020161179).

Two-year change in latent classes of comorbidity among high-risk Veterans in primary care: a brief report

Background

Segmentation models such as latent class analysis are an increasingly popular approach to inform group-tailored interventions for high-risk complex patients. Multiple studies have identified clinically meaningful high-risk segments, but few have evaluated change in groupings over time.

Objectives

To describe population-level and individual change over time in latent comorbidity groups among Veterans at high-risk of hospitalization in the Veterans Health Administration (VA).

Research design

Using a repeated cross-sectional design, we conducted a latent class analysis of chronic condition diagnoses. We compared latent class composition, patient high-risk status, and patient class assignment in 2018 to 2020.

Subjects

Two cohorts of eligible patients were selected: those active in VA primary care and in the top decile of predicted one-year hospitalization risk in 2018 (n = 951,771) or 2020 (n = 978,771).

Measures

Medical record data were observed from January 2016–December 2020. Latent classes were modeled using indicators for 26 chronic health conditions measured with a 2-year lookback period from study entry.

Results

Five groups were identified in both years, labeled based on high prevalence conditions: Cardiometabolic (23% in 2018), Mental Health (18%), Substance Use Disorders (16%), Low Diagnosis (25%), and High Complexity (10%). The remaining 8% of 2018 patients were not assigned to a group due to low predicted probability. Condition prevalence overall and within groups was stable between years. However, among the 563,725 patients identified as high risk in both years, 40.8% (n = 230,185) had a different group assignment in 2018 versus 2020.

Conclusions

In a repeated latent class analysis of nearly 1 million Veterans at high-risk for hospitalization, population-level groups were stable over two years, but individuals often moved between groups. Interventions tailored to latent groups need to account for change in patient status and group assignment over time.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12913-022-08757-x.

Clinical Outcome and Utilization Profiles Among Latent Groups of High-Risk Patients: Moving from Segmentation Towards Intervention

BACKGROUND

The ability of latent class models to identify clinically distinct groups among high-risk patients has been demonstrated, but it is unclear how healthcare data can inform group-specific intervention design.

OBJECTIVE

Examine how utilization patterns across latent groups of high-risk patients provide actionable information to guide group-specific intervention design.

DESIGN

Cohort study using data from 2012 to 2015.

PATIENTS

Participants were 934,787 patients receiving primary care in the Veterans Health Administration, with predicted probability of 12-month hospitalization in the top 10^th percentile during 2014.

MAIN MEASURES

Patients were assigned to latent groups via mixture-item response theory models based on 28 chronic conditions. We modeled odds of all-cause mortality, hospitalizations, and 30-day re-hospitalizations by group membership. Detailed outpatient and inpatient utilization patterns were compared between groups.

KEY RESULTS

A total of 764,257 (81.8%) of patients were matched with a comorbidity group. Groups were characterized by substance use disorders (14.0% of patients assigned), cardiometabolic conditions (25.7%), mental health conditions (17.6%), pain/arthritis (19.1%), cancer (15.3%), and liver disease (8.3%). One-year mortality ranged from 2.7% in the Mental Health group to 14.9% in the Cancer group, compared to 8.5% overall. In adjusted models, group assignment predicted significantly different odds of each outcome. Groups differed in their utilization of multiple types of care. For example, patients in the Pain group had the highest utilization of in-person primary care, with a mean (SD) of 5.3 (5.0) visits in the year of follow-up, while the Substance Use Disorder group had the lowest, with 3.9 (4.1) visits. The Substance Use Disorder group also had the highest rates of using services for housing instability (25.1%), followed by the Liver group (10.1%).

CONCLUSIONS

Latent groups of high-risk patients had distinct hospitalization and utilization profiles, despite having comparable levels of predicted baseline risk. Utilization profiles pointed towards system-specific care needs that could inform tailored interventions.

Empiric segmentation of high-risk patients: a structured literature review

OBJECTIVES

Empiric segmentation is a rapidly growing, learning health system approach that uses large health care system data sets to identify groups of high-risk patients who may benefit from similar interventions. We aimed to review studies that used data-driven approaches to segment high-risk patient populations and describe how their designs and findings can inform health care leaders who are interested in applying similar techniques to their patient populations.

STUDY DESIGN

Structured literature review.

METHODS

We searched for original research articles published since 2000 that identified high-risk adult patient populations and applied data-driven analyses to segment the population. Two reviewers independently extracted study population source and criteria for high-risk designation, segmentation method, data types included, model selection criteria, and model results from the identified studies.

RESULTS

Our search identified 224 articles, 12 of which met criteria for full review. Of these, 8 segmented high-risk patients and 4 segmented diagnoses without assigning patients to unique groups. Studies segmenting patients more often had clinically interpretable results. Common groups were defined by high prevalence of diabetes, cardiovascular disease, psychiatric conditions including substance use disorders, and neurologic disease (eg, stroke). Few studies incorporated patients' functional or social factors. Resulting patient and diagnosis clusters varied in ways closely linked to the model inputs, patient population inclusion criteria, and health care system context.

CONCLUSIONS

Empiric segmentation can yield clinically relevant groups of patients with complex medical needs. Segmentation results are context dependent, suggesting the need for careful design and interpretation of segmentation models to ensure that results can inform clinical care and program design in the target setting.

Outcomes of a randomized quality improvement trial for high-risk Veterans in year two

OBJECTIVE

The Veterans Health Administration (VHA) conducted a randomized quality improvement evaluation to determine whether augmenting patient-centered medical homes with Primary care Intensive Management (PIM) decreased utilization of acute care and health care costs among patients at high risk for hospitalization. PIM was cost-neutral in the first year; we analyzed changes in utilization and costs in the second year.

DATA SOURCES

VHA administrative data for five demonstration sites from August 2013 to March 2019.

DATA SOURCES

Administrative data extracted from VHA's Corporate Data Warehouse.

STUDY DESIGN

Veterans with a risk of 90-day hospitalization in the top 10th percentile and recent hospitalization or emergency department (ED) visit were randomly assigned to usual primary care vs primary care augmented by PIM. PIM included interdisciplinary teams, comprehensive patient assessment, intensive case management, and care coordination services. We compared the change in mean VHA inpatient and outpatient utilization and costs (including PIM expenses) per patient for the 12-month period before randomization and 13-24 months after randomization for PIM vs usual care using difference-in-differences.

PRINCIPAL FINDINGS

Both PIM patients (n = 1902) and usual care patients (n = 1882) had a mean of 5.6 chronic conditions. PIM patients had a greater number of primary care visits compared to those in usual care (mean 4.6 visits/patient/year vs 3.7 visits/patient/year, p < 0.05), but ED visits (p = 0.45) and hospitalizations (p = 0.95) were not significantly different. We found a small relative increase in outpatient costs among PIM patients compared to those in usual care (mean difference + $928/patient/year, p = 0.053), but no significant differences in mean inpatient costs (+$245/patient/year, p = 0.97). Total mean health care costs were similar between the two groups during the second year (mean difference + $1479/patient/year, p = 0.73).

CONCLUSIONS

Approaches that target patients solely based on the high risk of hospitalization are unlikely to reduce acute care use or total costs in VHA, which already offers patient-centered medical homes.

Subgroups of High-Risk Veterans Affairs Patients Based on Social Determinants of Health Predict Risk of Future Hospitalization

OBJECTIVE

Population segmentation has been recognized as a foundational step to help tailor interventions. Prior studies have predominantly identified subgroups based on diagnoses. In this study, we identify clinically coherent subgroups using social determinants of health (SDH) measures collected from Veterans at high risk of hospitalization or death.

STUDY DESIGN AND SETTING

SDH measures were obtained for 4684 Veterans at high risk of hospitalization through mail survey. Eleven self-report measures known to impact hospitalization and amenable to intervention were chosen a priori by the study team to identify subgroups through latent class analysis. Associations between subgroups and demographic and comorbidity characteristics were calculated through multinomial logistic regression. Odds of 180-day hospitalization were compared across subgroups through logistic regression.

RESULTS

Five subgroups of high-risk patients emerged-those with: minimal SDH vulnerabilities (8% hospitalized), poor/fair health with few SDH vulnerabilities (12% hospitalized), social isolation (10% hospitalized), multiple SDH vulnerabilities (12% hospitalized), and multiple SDH vulnerabilities without food or medication insecurity (10% hospitalized). In logistic regression, the "multiple SDH vulnerabilities" subgroup had greater odds of 180-day hospitalization than did the "minimal SDH vulnerabilities" reference subgroup (odds ratio: 1.53, 95% confidence interval: 1.09-2.14).

CONCLUSION

Self-reported SDH measures can identify meaningful subgroups that may be used to offer tailored interventions to reduce their risk of hospitalization and other adverse events.

Optimal Examination Sites for Periodontal Disease Evaluation: Applying the Item Response Theory Graded Response Model

Periodontal examination data have a complex structure. For epidemiological studies, mass screenings, and public health use, a simple index that represents the periodontal condition is necessary. Periodontal indices for partial examination of selected teeth have been developed. However, the selected teeth vary between indices, and a justification for the selection of examination teeth has not been presented. We applied a graded response model based on the item response theory to select optimal examination teeth and sites that represent periodontal conditions. Data were obtained from 254 patients who participated in a multicenter follow-up study. Baseline data were obtained from initial follow-up. Optimal examination sites were selected using item information calculated by graded response modeling. Twelve sites—maxillary 2nd premolar (palatal-medial), 1st premolar (palatal-distal), canine (palatal-medial), lateral incisor (palatal-central), central incisor (palatal-distal) and mandibular 1st premolar (lingual, medial)—were selected. Mean values for clinical attachment level, probing pocket depth, and bleeding on probing by full mouth examinations were used for objective variables. Measuring the clinical parameters of these sites can predict the results of full mouth examination. For calculating the periodontal index by partial oral examination, a justification for the selection of examination sites is essential. This study presents an evidence-based partial examination methodology and its modeling.

A Resampling Based Grid Search Method to Improve Reliability and Robustness of Mixture-Item Response Theory Models of Multimorbid High-Risk Patients

There are many statistics available to the applied statistician for assessing model fit and even more methods for assessing internal and external validity. We detail a useful approach using a grid search technique that balances the internal model consistency with generalizability and can be used with models that naturally lend themselves to multiple assessment techniques. Our method relies on resampling and a simple grid search method over 3 commonly used statistics that are simple to calculate. We apply this method in a latent traits framework using a mixture Item Response Theory (MIXIRT) model of common chronic health conditions. Model fit is assessed using Akaike's Information Criteria (AIC), latent class similarity is measured with the Variance of Information (VI), and the consistency of condition complexity and prevalence across latent classes is compared using Kendall's τ rank order statistic. From two patient cohorts at high risk for hospitalization in 2014 and 2018, we generated 19 MIXIRT models (allowing 2-20 latent classes) on 21 common comorbid conditions identified via healthcare encounter diagnosis codes. We ran these models on 100 bootstrap samples of size 10% for each cohort. Among the resulting models, combined AIC and VI statistics identified 5-7 latent classes, but the rank order correlation of condition complexity revealed that only the 5 class solutions had consistent condition complexity. The 5 class solutions were combined to produce a single parsimonious MIXIRT solution that balanced clinical significance with model fit, cluster similarity, and consistency of condition complexity.