Costas Kallis and Pete Dickson, Liverpool Data Hub
According to Public Health England, in the 2018-2019 period, there were 1.26 million hospital admissions in England related to alcohol consumption – 7.4% of all hospital admissions. Patients who have previously had an emergency alcohol-related admission to hospital are at increased risk of further admissions. Identifying patients at greatest risk of readmission would help primary care providers target interventions earlier, as identifying potential risk factors can reduce further harm to patients from alcohol and decrease the risk of readmissions. This collaborative project aimed to build a predictive model to understand the risk of emergency readmission within the next year for patients who had a previous alcohol-related emergency admission. This is helpful for commissioners to be able to plan services, predict bed use and to identify potential interventions to help mitigate unplanned care.
As part of the Connected Health Cities initiative, the Innovation Agency engaged the Hartree Centre team to work with clinical academics and statisticians at the University of Liverpool and business informatics teams at Liverpool CCG, to clean and combine anonymised data covering previous alcohol admissions in Liverpool. 6,209 potentially predictive inputs were generated. The team used statistical models to determine inputs with the most predictive value, working with clinicians to group into 28 clinically relevant features like age, underlying health conditions and comorbidities. These were added to a final model with 10 inputs and 71.2% accuracy. The threshold for flagging a positive prediction of readmission was adjusted to select a suitable number of patients according to service capacity.
The Hartree Centre used its expertise in applied machine learning to build models that selected the most relevant features from a vast array of diagnoses, patient history and demographic data. This collaborative approach with academic and clinical input allowed important features to be extracted and a live, working predictive model has been deployed within Liverpool CCG’s data warehouse. The tool will automatically flag high-risk patients earlier and can be used by local health services to help target interventions toward those most in need, ensuring access to the most appropriate services, thereby improving health outcomes for patients across the region. Reducing the need for patients to be readmitted would also provide long-term cost savings for local health care providers.
"A collaboration between clinicians and statisticians at the University of Liverpool, data scientists at Hartree Centre the Innovation Agency and business informatics teams at Liverpool CCG has delivered a product that can help patients with alcohol-related conditions be better managed. This project will pave the way for future collaborations to show machine learning being applied in improving patient outcomes."
This system could potentially be used in a variety of clinical settings such as:
- In the hospital, when a patient is discharged after an alcohol-related admission, to determine where further intervention is needed
- By specialist alcohol nurses, to target their efforts most appropriately
- In primary care, to identify patients who are at risk of alcohol-related admission and take appropriate steps to help them
- Applied to other clinical problems where a prediction of outcome is required, demonstrating the value of using machine learning techniques in a clinical setting