Identifying potential biases in code sequences in primary care electronic healthcare records: a retrospective cohort study of the determinants of code frequency

Comparing natural language processing representations of coded disease sequences for prediction in electronic health records

OBJECTIVE

Natural language processing (NLP) algorithms are increasingly being applied to obtain unsupervised representations of electronic health record (EHR) data, but their comparative performance at predicting clinical endpoints remains unclear. Our objective was to compare the performance of unsupervised representations of sequences of disease codes generated by bag-of-words versus sequence-based NLP algorithms at predicting clinically relevant outcomes.

MATERIALS AND METHODS

This cohort study used primary care EHRs from 6 286 233 people with Multiple Long-Term Conditions in England. For each patient, an unsupervised vector representation of their time-ordered sequences of diseases was generated using 2 input strategies (212 disease categories versus 9462 diagnostic codes) and different NLP algorithms (Latent Dirichlet Allocation, doc2vec, and 2 transformer models designed for EHRs). We also developed a transformer architecture, named EHR-BERT, incorporating sociodemographic information. We compared the performance of each of these representations (without fine-tuning) as inputs into a logistic classifier to predict 1-year mortality, healthcare use, and new disease diagnosis.

RESULTS

Patient representations generated by sequence-based algorithms performed consistently better than bag-of-words methods in predicting clinical endpoints, with the highest performance for EHR-BERT across all tasks, although the absolute improvement was small. Representations generated using disease categories perform similarly to those using diagnostic codes as inputs, suggesting models can equally manage smaller or larger vocabularies for prediction of these outcomes.

DISCUSSION AND CONCLUSION

Patient representations produced by sequence-based NLP algorithms from sequences of disease codes demonstrate improved predictive content for patient outcomes compared with representations generated by co-occurrence-based algorithms. This suggests transformer models may be useful for generating multi-purpose representations, even without fine-tuning.

Association between psoriasis and atrial fibrillation: A Systematic review and meta-analysis

INTRODUCTION

Psoriasis is a prevalent inflammatory skin condition characterized by erythematous plaques with scaling. Recent research has demonstrated an increased risk of cardiovascular diseases in patients with psoriasis; however, current evidence on atrial fibrillation (AF) risk in psoriasis is limited.

MATERIALS AND METHODS

A systematic literature search was performed on major bibliographic databases to retrieve studies that evaluated AF risk in patients with psoriasis. The DerSimonian and Laird random effects model was used to pool the hazard ratios (HR) with 95 % confidence intervals (CI). Subgroup analysis was conducted by dividing the patients into mild and severe psoriasis groups. Publication bias was assessed by visual inspection and Egger's regression test. Statistical significance was set at p < 0.05.

RESULTS

Seven studies were included, with 10,974,668 participants (1,94,230 in the psoriasis group and 10,780,439 in the control group). Patients with psoriasis had a significantly higher risk of AF [Pooled HR: 1.28; 95 % CI: 1.20, 1.36; p < 0.00001]. In subgroup analysis, patients with severe psoriasis [HR: 1.32; 95 % CI: 1.23, 1.42; p < 0.00001] demonstrated a slightly higher risk of AF, although statistically insignificant (p = 0.17), than the mild psoriasis group [HR: 1.21; 95 % CI: 1.10, 1.33; p < 0.0001]. Egger's regression test showed no statistically significant publication bias (p = 0.24).

CONCLUSION

Our analysis demonstrated that patients with psoriasis are at a significantly higher risk of AF and hence should be closely monitored for AF. Further large-scale and multicenter randomized trials are warranted to validate the robustness of our findings.

Identifying multi-resolution clusters of diseases in ten million patients with multimorbidity in primary care in England

BACKGROUND

Identifying clusters of diseases may aid understanding of shared aetiology, management of co-morbidities, and the discovery of new disease associations. Our study aims to identify disease clusters using a large set of long-term conditions and comparing methods that use the co-occurrence of diseases versus methods that use the sequence of disease development in a person over time.

METHODS

We use electronic health records from over ten million people with multimorbidity registered to primary care in England. First, we extract data-driven representations of 212 diseases from patient records employing (i) co-occurrence-based methods and (ii) sequence-based natural language processing methods. Second, we apply the graph-based Markov Multiscale Community Detection (MMCD) to identify clusters based on disease similarity at multiple resolutions. We evaluate the representations and clusters using a clinically curated set of 253 known disease association pairs, and qualitatively assess the interpretability of the clusters.

RESULTS

Both co-occurrence and sequence-based algorithms generate interpretable disease representations, with the best performance from the skip-gram algorithm. MMCD outperforms k-means and hierarchical clustering in explaining known disease associations. We find that diseases display an almost-hierarchical structure across resolutions from closely to more loosely similar co-occurrence patterns and identify interpretable clusters corresponding to both established and novel patterns.

CONCLUSIONS

Our method provides a tool for clustering diseases at different levels of resolution from co-occurrence patterns in high-dimensional electronic health records, which could be used to facilitate discovery of associations between diseases in the future.

Effect of timeframes to define long term conditions and sociodemographic factors on prevalence of multimorbidity using disease code frequency in primary care electronic health records: retrospective study

Objective

To determine the extent to which the choice of timeframe used to define a long term condition affects the prevalence of multimorbidity and whether this varies with sociodemographic factors.

Design

Retrospective study of disease code frequency in primary care electronic health records.

Data sources

Routinely collected, general practice, electronic health record data from the Clinical Practice Research Datalink Aurum were used.

Main outcome measures

Adults (≥18 years) in England who were registered in the database on 1 January 2020 were included. Multimorbidity was defined as the presence of two or more conditions from a set of 212 long term conditions. Multimorbidity prevalence was compared using five definitions. Any disease code recorded in the electronic health records for 212 conditions was used as the reference definition. Additionally, alternative definitions for 41 conditions requiring multiple codes (where a single disease code could indicate an acute condition) or a single code for the remaining 171 conditions were as follows: two codes at least three months apart; two codes at least 12 months apart; three codes within any 12 month period; and any code in the past 12 months. Mixed effects regression was used to calculate the expected change in multimorbidity status and number of long term conditions according to each definition and associations with patient age, gender, ethnic group, and socioeconomic deprivation.

Results

9 718 573 people were included in the study, of whom 7 183 662 (73.9%) met the definition of multimorbidity where a single code was sufficient to define a long term condition. Variation was substantial in the prevalence according to timeframe used, ranging from 41.4% (n=4 023 023) for three codes in any 12 month period, to 55.2% (n=5 366 285) for two codes at least three months apart. Younger people (eg, 50-75% probability for 18-29 years v 1-10% for ≥80 years), people of some minority ethnic groups (eg, people in the Other ethnic group had higher probability than the South Asian ethnic group), and people living in areas of lower socioeconomic deprivation were more likely to be re-classified as not multimorbid when using definitions requiring multiple codes.

Conclusions

Choice of timeframe to define long term conditions has a substantial effect on the prevalence of multimorbidity in this nationally representative sample. Different timeframes affect prevalence for some people more than others, highlighting the need to consider the impact of bias in the choice of method when defining multimorbidity.

Assigning disease clusters to people: A cohort study of the implications for understanding health outcomes in people with multiple long-term conditions

Background

Identifying clusters of co-occurring diseases may help characterise distinct phenotypes of Multiple Long-Term Conditions (MLTC). Understanding the associations of disease clusters with health-related outcomes requires a strategy to assign clusters to people, but it is unclear how the performance of strategies compare.

Aims

First, to compare the performance of methods of assigning disease clusters to people at explaining mortality, emergency department attendances and hospital admissions over one year. Second, to identify the extent of variation in the associations with each outcome between and within clusters.

Methods

We conducted a cohort study of primary care electronic health records in England, including adults with MLTC. Seven strategies were tested to assign patients to fifteen disease clusters representing 212 LTCs, identified from our previous work. We tested the performance of each strategy at explaining associations with the three outcomes over 1 year using logistic regression and compared to a strategy using the individual LTCs.

Results

6,286,233 patients with MLTC were included. Of the seven strategies tested, a strategy assigning the count of conditions within each cluster performed best at explaining all three outcomes but was inferior to using information on the individual LTCs. There was a larger range of effect sizes for the individual LTCs within the same cluster than there was between the clusters.

Conclusion

Strategies of assigning clusters of co-occurring diseases to people were less effective at explaining health-related outcomes than a person's individual diseases. Furthermore, clusters did not represent consistent relationships of the LTCs within them, which might limit their application in clinical research.