Digital health records and machine learning lead to decision support in hospitals

Ali Ebrahimi works as a postdoctoral researcher at SDU Health Informatics and Technology. He is a computer scientist, and his research area is clinical machine learning to develop predictive models for health-related problems.

Ali analysed digital health records to determine at an early stage Alcohol Use Disorder (AUD). AUD causes a significant amount of mortality and injuries. Most individuals suffering from AUD never undergo specialist treatment during their addiction due to the poor performance of conventional AUD identification methods, the lack of systematic screening for alcohol problems, and the taboo and stigma associated with harmful drinking. However, the recent availability of vast amounts of Electronic Health Records (EHR), which contain patients’ discharge data, and the advancement of Machine Learning (ML) algorithms have made it easier to offer clinical reasoning when making decisions. Therefore, the study used patients’ EHRs to develop predictive models to identify and detect patients with AUD. Ali and the other scientists involved collected the EHRs of 2,551 AUD-Positive and AUD-Negative patients from Odense University Hospital.

Anette Søgaard Nielsen looks at the benefits of the project for patients

Anette is Professor WSO and works at the Institute of Clinical Research at SDU. Together with colleagues, Anette Søgaard Nielsen researches in Psychiatry and Addiction Medicine.

Ali’s study is significant for her because the patients’ way of living affects their risk of illness and prognosis of treatment. That is particularly the case for smoking and alcohol use, both accounting for illnesses and preventable death. Therefore, hospital staff shall – in general – inform the patients about what they can do themselves to get healthy. This is very important if the patients are unaware that they have a habit that leads to or affects their illness or do not know where to seek help to change. In the latter situation, the staff shall advise and refer the patient to relevant help and further treatment. Hospital staff feel quite comfortable talking with the patient about the impact of smoking on health. Still, unfortunately, staff are much more reluctant to address the topic of alcohol due to fear of stigma.  Qualitative studies have shown that staff feel uncertain and wish to be rather sure that the patients’ alcohol use is problematic for their health before starting a conversation about alcohol and the benefits of reducing the alcohol intake. Therefore, patients are often not informed about how to prevent illness and rehospitalisation caused by alcohol use. They are not referred to treatment for alcohol use disorder, even if they need it.

All reduction of excessive alcohol use has a positive impact on health and other diseases. Late diagnostics lead to preventable death and severe illness, particularly cardiovascular diseases, liver diseases, and cancer.

About the collaboration

This project was a multidisciplinary project among computer engineers and clinical alcohol researchers. Numerous meetings were conducted to understand the nature of AUD, the availability of EHRs, and the anticipated outcomes. The meetings were necessary to establish trust and generate an understanding of the initial problem and its challenges during the development of the algorithms.

Moreover, all algorithms were tested using state-of-the-art evaluation and validation techniques. The final test will be, when clinical staff tries the algorithms in real situations.

The benefit of such a data-driven approach

A data-driven decision support tool can only help the staff become aware that excessive alcohol use may be a complicating factor and remind staff to have a conversation with the patient and offer help if needed. It is not a tool for giving the patient an additional diagnosis.

Our current developed models can detect AUD-Positive and AUD-Negative patients with an accuracy of 93% using Random Forest algorithm. On the other hand, we have developed models using Neural Networks algorithms for early detection of hazardous and harmful drinkers with an overall accuracy of 87% and 89% for hazardous and harmful drinkers for 18 months before AUD, respectively.

Colletion of data remains a difficulty

Ali sees the main pitfall that they faced in the data collection, which they solved by contacting the OUH data centre. It is necessary to be steady and follow up in such a process.

More interaction between clinics and researchers would probably enhance understanding problems and lead to more innovative solutions.

Leave a Reply

Your email address will not be published. Required fields are marked *