Using Classification and Regression Trees (CART) to Identify Prescribing Thresholds for Cardiovascular Disease

Predicting high health-cost users among people with cardiovascular disease using machine learning and nationwide linked social administrative datasets

OBJECTIVES

To optimise planning of public health services, the impact of high-cost users needs to be considered. However, most of the existing statistical models for costs do not include many clinical and social variables from administrative data that are associated with elevated health care resource use, and are increasingly available. This study aimed to use machine learning approaches and big data to predict high-cost users among people with cardiovascular disease (CVD).

METHODS

We used nationally representative linked datasets in New Zealand to predict CVD prevalent cases with the most expensive cost belonging to the top quintiles by cost. We compared the performance of four popular machine learning models (L1-regularised logistic regression, classification trees, k-nearest neighbourhood (KNN) and random forest) with the traditional regression models.

RESULTS

The machine learning models had far better accuracy in predicting high health-cost users compared with the logistic models. The harmony score F1 (combining sensitivity and positive predictive value) of the machine learning models ranged from 30.6% to 41.2% (compared with 8.6-9.1% for the logistic models). Previous health costs, income, age, chronic health conditions, deprivation, and receiving a social security benefit were among the most important predictors of the CVD high-cost users.

CONCLUSIONS

This study provides additional evidence that machine learning can be used as a tool together with big data in health economics for identification of new risk factors and prediction of high-cost users with CVD. As such, machine learning may potentially assist with health services planning and preventive measures to improve population health while potentially saving healthcare costs.

Predictive model to identify multiple failure to biological therapy in patients with rheumatoid arthritis

Background:

Despite advances in the treatment of rheumatoid arthritis (RA) and the wide range of therapies available, there is a percentage of patients whose treatment presents a challenge for clinicians due to lack of response to multiple biologic and target-specific disease-modifying antirheumatic drugs (b/tsDMARDs).

Objective:

To develop and validate an algorithm to predict multiple failure to biological therapy in patients with RA.

Design:

Observational retrospective study involving subjects from a cohort of patients with RA receiving b/tsDMARDs.

Methods:

Based on the number of prior failures to b/tsDMARDs, patients were classified as either multi-refractory (MR) or non-refractory (NR). Patient characteristics were considered in the statistical analysis to design the predictive model, selecting those variables with a predictive capability. A decision algorithm known as ‘classification and regression tree’ (CART) was developed to create a prediction model of multi-drug resistance. Performance of the prediction algorithm was evaluated in an external independent cohort using area under the curve (AUC).

Results:

A total of 136 patients were included: 51 MR and 85 NR. The CART model was able to predict multiple failures to b/tsDMARDs using disease activity score-28 (DAS-28) values at 6 months after the start time of the initial b/tsDMARD, as well as DAS-28 improvement in the first 6 months and baseline DAS-28. The CART model showed a capability to correctly classify 94.1% NR and 87.5% MR patients with a sensitivity = 0.88, a specificity = 0.94, and an AUC = 0.89 (95% CI: 0.74–1.00). In the external validation cohort, 35 MR and 47 NR patients were included. The AUC value for the CART model in this cohort was 0.82 (95% CI: 0.73–0.9).

Conclusion:

Our model correctly classified NR and MR patients based on simple measurements available in routine clinical practice, which provides the possibility to characterize and individualize patient treatments during early stages.

An examination of machine learning to map non-preference based patient reported outcome measures to health state utility values

Non-preference-based patient-reported outcome measures (PROMs) are popular in health outcomes research. These measures, however, cannot be used to estimate health state utilities, limiting their usefulness for economic evaluations. Mapping PROMs to a multi-attribute utility instrument is one solution. While mapping is commonly conducted using econometric techniques, failing to specify the complex interactions between variables may lead to inaccurate prediction of utilities, resulting in inaccurate estimates of cost-effectiveness and suboptimal funding decisions. These issues can be addressed using machine learning. This paper evaluates the use of machine learning as a mapping tool. We adopt a comprehensive approach to compare six machine learning techniques with eight econometric techniques to map the Patient-Reported Outcomes Measurement Information System Global Health 10 (PROMIS-GH10) to the EuroQol five dimensions (EQ-5D-5L). Using data collected from 2015 Australians, we find the least absolute shrinkage and selection operator (LASSO) model out-performed all machine learning techniques and the adjusted limited dependent variable mixture model (ALDVMM) out-performed all econometric techniques, with the LASSO performing better than ALDVMM. The variable selection feature of LASSO was then used to enhance the performance of the ALDVMM in a hybrid model. Our analysis identifies the potential benefits and challenges of using machine learning techniques for mapping and offers important insights for future research.

Haplotype heterogeneity and low linkage disequilibrium reduce reliable prediction of genotypes for the ‑α 3.7I form of α-thalassaemia using genome-wide microarray data

Background: The -α ^3.7I-thalassaemia deletion is very common throughout Africa because it protects against malaria. When undertaking studies to investigate human genetic adaptations to malaria or other diseases, it is important to account for any confounding effects of α-thalassaemia to rule out spurious associations.

Methods: In this study, we have used direct α-thalassaemia genotyping to understand why GWAS data from a large malaria association study in Kilifi Kenya did not identify the α-thalassaemia signal. We then explored the potential use of a number of new approaches to using GWAS data for imputing α-thalassaemia as an alternative to direct genotyping by PCR.

Results: We found very low linkage-disequilibrium of the directly typed data with the GWAS SNP markers around α-thalassaemia and across the haemoglobin-alpha ( HBA) gene region, which along with a complex haplotype structure, could explain the lack of an association signal from the GWAS SNP data. Some indirect typing methods gave results that were in broad agreement with those derived from direct genotyping and could identify an association signal, but none were sufficiently accurate to allow correct interpretation compared with direct typing, leading to confusing or erroneous results.

Conclusions: We conclude that going forwards, direct typing methods such as PCR will still be required to account for α-thalassaemia in GWAS studies.

Does coaching matter? Examining the impact of specific practice facilitation strategies on implementation of quality improvement interventions in the Healthy Hearts in the Heartland study

BACKGROUND

Practice facilitation is a multicomponent implementation strategy used to improve the capacity for practices to address care quality and implementation gaps. We sought to assess whether practice facilitators use of coaching strategies aimed at improving self-sufficiency were associated with improved implementation of quality improvement (QI) interventions in the Healthy Hearts in the Heartland Study.

METHODS

We mapped 27 practice facilitation activities to a framework that classifies practice facilitation strategies by the degree to which the practice develops its own process expertise (Doing Tasks, Project Management, Consulting, Teaching, and Coaching) and then used regression tree analysis to group practices by facilitation strategies experienced. Kruskal-Wallis tests were used to assess whether practice groups identified by regression tree analysis were associated with successful implementation of QI interventions and practice and study context variables.

RESULTS

There was no association between number of strategies performed by practice facilitators and number of QI interventions implemented. Regression tree analysis identified 4 distinct practice groups based on the number of Project Management and Coaching strategies performed. The median number of interventions increased across the groups. Practices receiving > 4 project management and > 6 coaching activities implemented a median of 17 of 35 interventions. Groups did not differ significantly by practice size, association with a healthcare network, or practice type. Statistically significant differences in practice location, number and duration of facilitator visits, and early study termination emerged among the groups, compared to the overall practice population.

CONCLUSIONS

Practices that engage in more coaching-based strategies with practice facilitators are more likely to implement more QI interventions, and practice receptivity to these strategies was not dependent on basic practice demographics.

Haplotype heterogeneity and low linkage disequilibrium reduce reliable prediction of genotypes for the ‑α3.7I form of α-thalassaemia using genome-wide microarray data

Background: The -α3.7I-thalassaemia deletion is very common throughout Africa because it protects against malaria. When undertaking studies to investigate human genetic adaptations to malaria or other diseases, it is important to account for any confounding effects of α-thalassaemia to rule out spurious associations. Methods: In this study we have used direct α-thalassaemia genotyping to understand why GWAS data from a large malaria association study in Kilifi Kenya did not identify the α-thalassaemia signal. We then explored the potential use of a number of new approaches to using GWAS data for imputing α-thalassaemia as an alternative to direct genotyping by PCR. Results: We found very low linkage-disequilibrium of the directly typed data with the GWAS SNP markers around α-thalassaemia and across the haemoglobin-alpha (HBA) gene region, which along with a complex haplotype structure, could explain the lack of an association signal from the GWAS SNP data. Some indirect typing methods gave results that were in broad agreement with those derived from direct genotyping and could identify an association signal, but none were sufficiently accurate to allow correct interpretation compared with direct typing, leading to confusing or erroneous results. Conclusions: We conclude that going forwards, direct typing methods such as PCR will still be required to account for α-thalassaemia in GWAS studies.

Pharmacological treatment of increased vascular risk and cognitive performance in middle-aged and old persons: six-year observational longitudinal study

Background

Lowering vascular risk is associated with a decrease in the prevalence of cardiovascular disease and dementia. However, it is still unknown whether lowering of vascular risk with pharmacological treatment preserves cognitive performance in general. Therefore, we compared the change in cognitive performance in persons with and without treatment of vascular risk factors.

Methods

In this longitudinal observational study, 256 persons (mean age, 58 years) were treated for increased vascular risk during a mean follow-up period of 5.5 years (treatment group), whereas 1678 persons (mean age, 50 years) did not receive treatment (control group). Cognitive performance was three times measured during follow-up using the Ruff Figural Fluency Test (RFFT) and Visual Association Test (VAT), and calculated as the average of standardized RFFT and VAT score per participant. Because treatment allocation was nonrandomized, additional analyses were performed in demographic and vascular risk-matched samples and adjusted for propensity scores.

Results

In the treatment group, mean (SD) cognitive performance changed from − 0.30 (0.80) to − 0.23 (0.80) to 0.02 (0.87), and in control group, from 0.08 (0.77) to 0.24 (0.79) to 0.49 (0.74) at the first, second and third measurement, respectively (p_trend < 0.001). After adjustment for demographics and vascular risk, the change in cognitive performance during follow-up was not statistically significantly different between the treatment and control group: mean estimated difference, − 0.10 (95%CI − 0.21 to 0.01; p = 0.08). Similar results were found in matched samples and after adjustment for propensity score.

Conclusion

Change in cognitive performance during follow-up was similar in treated and untreated persons. This suggests that lowering vascular risk preserves cognitive performance.

An exploratory model for the non-fatal drowning risks in children in Guangdong, China

Background

Drowning is a leading cause of accidental death in children under 14 years of age in Guangdong, China. We developed a statistical model to classify the risk of drowning among children based on the risk factors.

Methods

A multiple-stage cluster random sampling was employed to select the students in Grades 3 to 9 in two townships in Qingyuan, Guangdong. Questionnaire was a self-reported measure consisting of general information, knowledge, attitudes and activities. A univariate logistic regression model was used to preliminarily select the independent variables at a P value of 0.1 for multivariable model. Three-quarters of the participants were randomly selected as a training sample to establish the model, and the remaining were treated as a testing sample to validate the model.

Results

A total of 8390 children were included in this study, about 12.18% (1013) experienced drowning during the past one year. In the univariate logistic regression model, introvert personality, unclear distributions of water areas on the way to school, and bad relationships with their classmates and families were positively associated with drowning. However, females, older age and lower swimming skills were negatively associated with drowning. After employing the prediction model with these factors to estimate drowning risk of the students in the testing samples, the results of Hosmer-Lemeshow tests showed non-significant differences between the predictive results and actual risk (χ2 = 5.97, P = 0.65).

Conclusions

Male, younger children, higher swimming skills, bad relationship with their classmates and families, introvert personality and unclear distributions of water areas on the way to school were important risk factors of non-fatal drowning among children. The prediction model based on these variables has an acceptable predictive ability.

Cross-sectional analysis of the prevalence and predictors of statin utilisation in Ireland with a focus on primary prevention of cardiovascular disease

OBJECTIVE

To describe the prevalence of statin utilisation by people aged over 50 years in Ireland and the factors associated with the likelihood of using a statin, focusing particularly on those using statins for primary prevention of cardiovascular disease (CVD).

METHODS

This is a cross-sectional analysis of cardiovascular risk and sociodemographic factors associated with statin utilisation from wave 1 of The Irish Longitudinal Study on Ageing. A hierarchy of indications for statin utilisation, consisting of eight mutually exclusive levels of CVD-related diagnoses, was created. Participants were assigned one level of indication. The prevalence of statin utilisation was calculated. The likelihood that an individual was using a statin was estimated using a multivariable logistic regression model, controlling for cardiovascular risk and sociodemographic factors.

RESULTS

In this nationally representative sample (n=5618) of community-dwelling participants aged 50 years and over, 1715 (30.5%) were taking statins. Of these, 65.0% (57.3% of men and 72.7% of women) were doing so for the primary prevention of CVD. Thus, almost two-thirds of those taking statins did so for primary prevention and there was a notable difference between women and men in this regard. We also found that statin utilisation was highest among those with a prior history of CVD and was significantly associated with age (compared with the base category 50-64 years; 65-74 years OR 1.38 (95% CI 1.16 to 1.65); 75+ OR 1.33 (95% CI 1.04 to 1.69)), living with a spouse or partner (compared with the base category living alone; OR 1.35 (95% CI 1.10 to 1.65)), polypharmacy (OR 1.74 (95% CI 1.39 to 2.19)) and frequency of general practitioner visits (compared with the base category 0 visits per year; 1-2 visits OR 2.46 (95% CI 1.80 to 3.35); 3-4 visits OR 3.24 (95% CI 2.34 to 4.47); 5-6 visits OR 2.98 (95% CI 2.08 to 4.26); 7+ visits OR 2.51 (95% CI 1.73 to 3.63)), even after controlling for clinical need. There was no association between using statins and gender, education, income, social class, health insurance status, location or Systematic Coronary Risk Evaluation (SCORE) risk in the multivariable analysis.

CONCLUSION

Statin utilisation among those with no history of CVD accounted for almost two-thirds of all statin use, in part reflecting the high proportion of the population with no history of CVD, although utilisation rates were highest among those with a history of CVD.

Australian general practitioners initiate statin therapy primarily on the basis of lipid levels; New Zealand general practitioners use absolute risk

OBJECTIVES

To compare the determinants of initial statin prescribing between New Zealand and Australia. New Zealand has a system-wide absolute risk-based approach to primary care cardiovascular disease (CVD) management, while Australia has multiple guidelines.

METHOD

Classification and Regression Tree (CART) analysis of two observational studies of primary care CVD management from New Zealand (PREDICT-CVD) and Australia (AusHeart). Over 80% of eligible New Zealanders have been screened for CVD risk. PREDICT-CVD is used by approximately one-third of New Zealand GPs to perform web-based CVD risk assessment in routine practice, with the sample consisting of 126,519 individuals risk assessed between 1 January 2007 and 30 June 2014. AusHeart is a cluster-stratified survey of primary care CVD management that enrolled 534 GPs from across Australia, who in turn recruited 1381 patients between 1 April and 30 June 2008. Eligibility was restricted to 55-74year old patients without prior CVD.

RESULTS

The CART analyses demonstrated that New Zealand GPs prescribe statins primarily on the basis of absolute risk, while their Australian counterparts are influenced by a variety of individual risk factors, including total cholesterol, LDL cholesterol and diabetes.

CONCLUSIONS

Countries seeking to improve their management of CVD should consider adopting a 'whole of system' absolute risk-based approach with clear guidelines that are consistent with drug reimbursement rules; and include computerized decision-support tools that aid decision-making and allow monitoring of outcomes and continual improvement of practice.

Prognostic factors analysis in EGFR mutation-positive non-small cell lung cancer with brain metastases treated with whole brain-radiotherapy and EGFR-tyrosine kinase inhibitors

The survival time of non-small cell lung cancer (NSCLC) patients with brain metastases has been previously reported to be 6.5-10.0 months, even with systematic treatment. Patients that possess a certain epidermal growth factor receptor (EGFR) mutation alongside NSCLC with brain metastases also have a short survival rate, and a reliable prognostic model for these patients demonstrates a strong correlation between the outcome and treatment recommendations. The Cox proportional hazards regression and classification tree models were used to explore the prognostic factors in EGFR mutation-positive NSCLC patients with brain metastases following whole-brain radiation therapy (WBRT) and EGFR-tyrosine kinase inhibitor (EGFR-TKI) treatment. A total of 66 EGFR mutation-positive NSCLC patients with brain metastases were retrospectively reviewed. Univariate and multivariate analyses by Cox proportional hazards regression were then performed. The classification tree model was applied in order to identify prognostic groups of the patients. In the survival analysis, age, carcinoembryonic antigen (CEA) and status of the primary tumor were prognostic factors for progression free survival (P=0.006, 0.014 and 0.005, respectively) and overall survival (P=0.009, 0.013 and 0.009, respectively). The classification tree model was subsequently applied, which revealed 3 patient groups with significantly different survival times: Group I, age <65 years and CEA ≤10 µg/ml; Group II, age <65 years and CEA >10 µg/ml or age ≥65 years and CEA ≤10 µg/ml; and Group III, age ≥65 years and CEA >10 µg/ml. The major prognostic predictors for EGFR mutation-positive NSCLC patients with brain metastases following WBRT and EGFR-TKI were age and CEA. In addition, primary tumor control may be important for predicting survival.