Dataset Information

Derivation and validation of a chief complaint shortlist for unscheduled acute and emergency care in Uganda.

ABSTRACT:

Objectives

Derive and validate a shortlist of chief complaints to describe unscheduled acute and emergency care in Uganda.

Setting

A single, private, not-for profit hospital in rural, southwestern Uganda.

Participants

From 2009 to 2015, 26 996 patient visits produced 42 566 total chief complaints for the derivation dataset, and from 2015 to 2017, 10 068 visits produced 20 165 total chief complaints for the validation dataset.

Methods

A retrospective review of an emergency centre quality assurance database was performed. Data were abstracted, cleaned and refined using language processing in Stata to produce a longlist of chief complaints, which was collapsed via a consensus process to produce a shortlist and turned into a web-based tool. This tool was used by two local Ugandan emergency care practitioners to categorise complaints from a second longlist produced from a separate validation dataset from the same study site. Their agreement on grouping was analysed using Cohen's kappa to determine inter-rater reliability. The chief complaints describing 80% of patient visits from automated and consensus shortlists were combined to form a candidate chief complaint shortlist.

Results

Automated data cleaning and refining recognised 95.8% of all complaints and produced a longlist of 555 chief complaints. The consensus process yielded a shortlist of 83 grouped chief complaints. The second validation dataset was reduced in Stata to a longlist of 451 complaints. Using the shortlist tool to categorise complaints produced 71.5% agreement, yielding a kappa of 0.70 showing substantial inter-rater reliability. Only one complaint did not fit into the shortlist and required a free-text amendment. The two shortlists were identical for the most common 14 complaints and combined to form a candidate list of 24 complaints that could characterise over 80% of all emergency centre chief complaints.

Conclusions

Shortlists of chief complaints can be generated to improve standardisation of data entry, facilitate research efforts and be employed for paper chart usage.

SUBMITTER: Rice BT

PROVIDER: S-EPMC6020949 | biostudies-literature |

REPOSITORIES: biostudies-literature

ACCESS DATA

Similar Datasets

Project description:Syndromic surveillance detects and monitors individual and population health indicators through sources such as emergency department records. Automated classification of these records can improve outbreak detection speed and diagnosis accuracy. Current syndromic systems rely on hand-coded keyword-based methods to parse written fields and may benefit from the use of modern supervised-learning classifier models. In this paper, we implement two recurrent neural network models based on long short-term memory (LSTM) and gated recurrent unit (GRU) cells and compare them to two traditional bag-of-words classifiers: multinomial naïve Bayes (MNB) and a support vector machine (SVM). The MNB classifier is one of only two machine learning algorithms currently being used for syndromic surveillance. All four models are trained to predict diagnostic code groups as defined by Clinical Classification Software, first to predict from discharge diagnosis, and then from chief complaint fields. The classifiers are trained on 3.6 million de-identified emergency department records from a single United States jurisdiction. We compare performance of these models primarily using the F1 score, and we measure absolute model performance to determine which conditions are the most amenable to surveillance based on chief complaint alone. Using discharge diagnoses, the LSTM classifier performs best, though all models exhibit an F1 score above 96.00. Using chief complaints, the GRU performs best (F1 = 47.38), and MNB with bigrams performs worst (F1 = 39.40). We also note that certain syndrome types are easier to detect than others. For example, chief complaints using the GRU model predicts alcohol-related disorders well (F1 = 78.91) but predicts influenza poorly (F1 = 14.80). In all instances, the RNN models outperformed the bag-of-words classifiers suggesting deep learning models could substantially improve the automatic classification of unstructured text for syndromic surveillance.

Project description:ObjectiveTo develop and externally validate a risk algorithm (QAdmissions) to estimate the risk of emergency hospital admission for patients aged 18-100 years in primary care.DesignProspective open cohort study using routinely collected data from general practice linked to hospital episode data during the 2-year study period 1 January 2010 to 31 December 2011.Setting405 general practices in England contributing to the national QResearch database to develop the algorithm. Two validation cohorts to validate the algorithm (1) 202 different QResearch practices and (2) 343 practices in England contributing to the Clinical Practice Research DataLink (CPRD). All general practices had data linked to hospital episode statistics at the individual patient level.ParticipantsWe studied 2 849 381 patients aged 18-100 years in the derivation cohort with over 4.6 million person-years of follow-up. 265 573 of these patients had one or more emergency admissions during follow-up. For the QResearch validation cohort, we identified 1 340 622 patients aged 18-100 years with over 2.2 million person-years of follow-up. Of these patients, 132 723 had one or more emergency admissions during follow-up. The CPRD cohort included 2 475 360 patients aged 18-100 years with over 3.8 million person-years of follow-up. 234 204 of these patients had one or more emergency admissions during follow-up. We excluded patients without a valid NHS number and a valid Townsend score.EndpointFirst (ie, incident) emergency admission to hospital in the next 2 years as recorded on the linked hospital episodes records.Risk factorsCandidate variables recorded on the general practitioner computer system including (1) demographic variables (age, sex, strategic health authority, Townsend deprivation score, ethnicity); (2) lifestyle variables (smoking, alcohol intake); (3) chronic diseases; (4) prescribed medication; (5) clinical values (body mass index, systolic blood pressure); (6) laboratory test results (haemoglobin, platelets, erythrocyte sedimentation rate, ratio of total serum cholesterol to high density lipoprotein cholesterol concentrations, liver function tests). We also included the number of emergency admissions in the preceding year based on information recorded on the linked hospital episodes records.ResultsThe final QAdmissions algorithm incorporated 30 variables. When applied to the QResearch validation cohort, it explained 41% of the variation in women and 43% of that in men. The D statistic for QAdmissions was 1.7 in women and 1.8 in men. The receiver operating curve statistic was 0.78 for men and 0.77 for women. QAdmissions had good performance on all measures of discrimination and calibration. The positive predictive value for emergency admissions for the top tenth of patients at highest risk was 42% and the sensitivity was 39%. The results for the CPRD validation cohort were similar.ConclusionsThe QAdmissions model provided a valid measure of absolute risk of emergency admission to hospital in the general population as shown by its performance in a separate validation cohort. Further research is needed to evaluate the cost-effectiveness of using these algorithms in primary care.