# Using machine learning to identify frequent attendance at accident and emergency services in Lanarkshire

## TL;DR

This study uses machine learning to predict frequent visits to emergency services in Lanarkshire and identifies key risk factors for such attendance.

## Contribution

The study evaluates multiple machine learning models and their effectiveness in predicting frequent A&E attendance in a specific healthcare region.

## Key findings

- The multi-layer perceptron model achieved the highest F1 score (0.75) for predicting frequent A&E attendance.
- Key health conditions and risk factors were consistently identified across models, though with some dataset-specific variations.
- Combining machine learning models improved prediction accuracy and provided insights into risk factors for frequent attendance.

## Abstract

Frequent attenders to accident and emergency (A&E) services pose complex challenges for healthcare providers, often driven by critical clinical needs. Machine learning (ML) offers potential for predictive approaches to managing frequent attendance, yet its application in this area is limited. Existing studies often focus on specific populations or models, raising concerns about generalisability. Identifying risk factors for frequent attendance and high resource use is crucial for effective prevention strategies.

This research aims to evaluate the strengths and weaknesses of ML approaches in predicting frequent A&E attendance in NHS Lanarkshire, Scotland, identify associated risk factors and compare findings with existing research to uncover commonalities and differences.

Health and social care data were collected from 17,437 A&E patients in NHS Lanarkshire (2021–2022), including clinical, social and demographic information. Five classification models were tested: multinomial logistic regression (LR), random forests (RF), support vector machine (SVM) classifier, k-nearest neighbours (k-NN) and multi-layer perceptron (MLP) classifier. Models were evaluated using a confusion matrix and metrics such as precision, recall, F1 and area under the curve. Shapley values were used to identify risk factors.

MLP achieved the highest F1 score (0.75), followed by k-NN, RF and SVM (0.72 each), and LR (0.70). Key health conditions and risk factors consistently predicted frequent attendance across models, with some variation highlighting dataset-specific characteristics.

This study underscores the utility of combining ML models to enhance prediction accuracy and identify risk factors. Findings align with existing research but reveal unique insights specific to the dataset and methodology.