Machine Learning for Feature Selection and Cluster Analysis in Drug Utilisation Research

Classification of patients with osteoarthritis through clusters of comorbidities using 633 330 individuals from Spain

OBJECTIVES

To explore clustering of comorbidities among patients with a new diagnosis of OA and estimate the 10-year mortality risk for each identified cluster.

METHODS

This is a population-based cohort study of individuals with first incident diagnosis of OA of the hip, knee, ankle/foot, wrist/hand or 'unspecified' site between 2006 and 2020, using SIDIAP (a primary care database representative of Catalonia, Spain). At the time of OA diagnosis, conditions associated with OA in the literature that were found in ≥1% of the individuals (n = 35) were fitted into two cluster algorithms, k-means and latent class analysis. Models were assessed using a range of internal and external evaluation procedures. Mortality risk of the obtained clusters was assessed by survival analysis using Cox proportional hazards.

RESULTS

We identified 633 330 patients with a diagnosis of OA. Our proposed best solution used latent class analysis to identify four clusters: 'low-morbidity' (relatively low number of comorbidities), 'back/neck pain plus mental health', 'metabolic syndrome' and 'multimorbidity' (higher prevalence of all studied comorbidities). Compared with the 'low-morbidity' cluster, the 'multimorbidity' cluster had the highest risk of 10-year mortality (adjusted hazard ratio [HR]: 2.19 [95% CI: 2.15, 2.23]), followed by the 'metabolic syndrome' cluster (adjusted HR: 1.24 [95% CI: 1.22, 1.27]) and the 'back/neck pain plus mental health' cluster (adjusted HR: 1.12 [95% CI: 1.09, 1.15]).

CONCLUSION

Patients with a new diagnosis of OA can be clustered into groups based on their comorbidity profile, with significant differences in 10-year mortality risk. Further research is required to understand the interplay between OA and particular comorbidity groups, and the clinical significance of such results.

K-means clustering of outpatient prescription claims for health insureds in Iran

OBJECTIVE

The segmentation of consumers based on their behavior and needs is the most crucial action of the health insurance organization. This study's objective is to cluster Iranian health insureds according to their demographics and data on outpatient prescriptions.

SETTING

The population in this study corresponded to the research sample. The Health Insurance Organization's outpatient claims were registered consecutively in 2016, 2017, 2018, and 2019 were clustered.

DESIGN

The k-means clustering algorithm was used to cross-sectionally and retrospectively analyze secondary data from outpatient prescription claims for secondary care using Python 3.10.

PARTICIPANTS

The current analysis transformed 21 776 350 outpatient prescription claims from health insured into 193 552 insureds.

RESULTS

Insureds using IQR were split into three classes: low, middle, and high risk. Based on the silhouette coefficient, the insureds of all classes were divided into three clusters. In all data for a period of four years, the first through third clusters, there were 21 799, 7170, and 19 419 insureds in the low-risk class. Middle-risk class had 48 348,23 321, 25 107 insureds, and 14 037, 28 504, 5847 insured in the high-risk class were included. For the first cluster of low-risk insureds: the total average cost of prescriptions paid by the insurance for the insureds was $211, the average age was 26 years, the average franchise was 88.5US$, the average number of medications and prescriptions were 409 and 62, the total average costs of prescriptions Outpatient was 302.5 US$, the total average number of medications for acute and chronic disease was 178 and 215, respectively. The majority of insureds were men, and those who were part of the householder's family.

CONCLUSIONS

By segmenting insurance customers, insurers can set insurance premium rates, controlling the risk of loss, which improves their capacity to compete in the insurance market.

Development of an effective clustering algorithm for older fallers

Falls are common and often lead to serious physical and psychological consequences for older persons. The occurrence of falls are usually attributed to the interaction between multiple risk factors. The clinical evaluation of falls risks is time-consuming as a result, hence limiting its availability. The purpose of this study was, therefore, to develop a clustering-based algorithm to determine falls risk. Data from the Malaysian Elders Longitudinal Research (MELoR), comprising 1411 subjects aged ≥55 years, were utilized. The proposed algorithm was developed through the stages of: data pre-processing, feature identification and extraction with either t-Distributed Stochastic Neighbour Embedding (t-SNE) or principal component analysis (PCA)), clustering (K-means clustering, Hierarchical clustering, and Fuzzy C-means clustering) and characteristics interpretation with statistical analysis. A total of 1279 subjects and 9 variables were selected for clustering after the data pre-possessing stage. Using feature extraction with the t-SNE and the K-means clustering algorithm, subjects were clustered into low, intermediate A, intermediate B and high fall risk groups which corresponded with fall occurrence of 13%, 19%, 21% and 31% respectively. Slower gait, poorer balance, weaker muscle strength, presence of cardiovascular disorder, poorer cognitive performance, and advancing age were the key variables identified. The proposed fall risk clustering algorithm grouped the subjects according to features. Such a tool could serve as a case identification or clinical decision support tool for clinical practice to enhance access to falls prevention efforts.

Consensus clustering for case series identification and adverse event profiles in pharmacovigilance

OBJECTIVE

To describe and evaluate vigiGroup - a consensus clustering algorithm which can identify groups of individual case reports referring to similar suspected adverse drug reactions and describe associated adverse event profiles, accounting for co-reported adverse event terms.

MATERIALS AND METHODS

Consensus clustering is achieved by grouping pairs of reports that are repeatedly placed together in the same clusters across a set of mixture model-based cluster analyses. The latter use empirical Bayes statistical shrinkage for improved performance. As baseline comparison, we considered a regular mixture model-based cluster analysis. Three randomly selected drugs in VigiBase, the World Health Organization's global database of Individual Case Safety Reports were analyzed: sumatriptan, ambroxol and tacrolimus. Clustering stability was assessed using the adjusted Rand index, ranging between -1 and +1, and clinical coherence was assessed through an intruder detection analysis.

RESULTS

For the three drugs considered, vigiGroup achieved stable and coherent results with adjusted Rand indices between +0.80 and +0.92, and intruder detection rates between 86% and 94%. Consensus clustering improved both stability and clinical coherence compared to mixture model-based clustering alone. Statistical shrinkage improved the stability of clusters compared to the baseline mixture model, as well as the cross-validated log-likelihood.

CONCLUSIONS

The proposed algorithm can achieve adequate stability and clinical coherence in clustering individual case reports, thereby enabling better identification of case series and associated adverse event profiles in pharmacovigilance. The use of empirical Bayes shrinkage and consensus clustering each led to meaningful improvements in performance.

Exploring patient multimorbidity and complexity using health insurance claims data: a cluster analysis approach (Preprint)

Background

Although the trend of progressing morbidity is widely recognized, there are numerous challenges when studying multimorbidity and patient complexity. For multimorbid or complex patients, prone to fragmented care and high health care use, novel estimation approaches need to be developed.

Objective

This study aims to investigate the patient multimorbidity and complexity of Swiss residents aged ≥50 years using clustering methodology in claims data.

Methods

We adopted a clustering methodology based on random forests and used 34 pharmacy-based cost groups as the only input feature for the procedure. To detect clusters, we applied hierarchical density-based spatial clustering of applications with noise. The reasonable hyperparameters were chosen based on various metrics embedded in the algorithms (out-of-bag misclassification error, normalized stress, and cluster persistence) and the clinical relevance of the obtained clusters.

Results

Based on cluster analysis output for 18,732 individuals, we identified an outlier group and 7 clusters: individuals without diseases, patients with only hypertension-related diseases, patients with only mental diseases, complex high-cost high-need patients, slightly complex patients with inexpensive low-severity pharmacy-based cost groups, patients with 1 costly disease, and older high-risk patients.

Conclusions

Our study demonstrated that cluster analysis based on pharmacy-based cost group information from claims-based data is feasible and highlights clinically relevant clusters. Such an approach allows expanding the understanding of multimorbidity beyond simple disease counts and can identify the population profiles with increased health care use and costs. This study may foster the development of integrated and coordinated care, which is high on the agenda in policy making, care planning, and delivery.