Can we anticipate use of social and health care services with machine learning?

Effective utilisation of data accumulated in customer and patient information systems is key to new proactive social and health care practices. Elderly people, whose share of social welfare and health care expenditure is many times higher than the population average, are a particularly interesting target group. To predict the multidisciplinary service use of the elderly, the MAITE project studied how artificial intelligence could be applied.
Lue tämä suomeksi

Photo by Irwan @blogcious on Unsplash

The key idea was to identify groups of people with a higher-than-average risk of abundant, multidisciplinary use of social welfare and health care services. The joint project was led by Finnish Institute for Health and Welfare THL, with partners the University of Lapland, University of Helsinki, VTT and Päijät-Häme wellbeing services county. VTT's task in the project was to develop and demonstrate a prototype of a prediction model based on artificial intelligence.

Several previous projects have developed AI-based models to support early diagnosis of specific diseases or disease groups or to predict the course of disease. In the MAITE project, prediction was not targeted specifically at any illness, but at the use of social welfare and health care services in general. It was essential to identify groups of people who use these services more often and more diversely than others. The project extensively examined the process of developing the prediction model, aiming to increase the capacity to implement and introduce AI-based solutions in the sector.

Prediction model developed step by step

The objectives for the prediction model and the required information from the Päijät-Häme customer and patient registers were defined together with the project partners. The Päijät-Häme wellbeing services county granted permission to use the pseudonymised data in scientific research in accordance with the act on the secondary use of health and social data.

The data was processed to develop the machine learning model in Kapseli, Findata's secure operating environment. The ICT team of Päijät-Häme wellbeing services county was responsible for the extraction of the material from the organisation's database and for the transfer of the material to Kapseli. VTT implemented the data pre-processing, feature selection, model training, validation and testing for the machine learning model.

After the developed machine learning model was considered anonymous, it was deployed in VTT's demo environment, where it can be studied, for example, by illustrating the impact of different factors on the risk of unwanted service use. Additionally, the characteristics of the model data are visualised with an interactive geographic color-coded map, which allows the user to familiarise themselves with the main features of the data set used to develop the model.

The social welfare and health care actors wanted to forecast the use of services that would not be necessary or that could be prevented by proactive interventions. It was considered difficult to meet this wish because the customer and patient data on which the forecast is based do not provide information on whether the use of the services was necessary or preventable. Instead, the number of acute visits and the number of different social welfare and health care services used concurrently were utilized as key indicators for undesired service use.

Model forecasts customer's risk for unwanted service use

The implemented prediction model forecasts whether a customer is at risk of using social welfare and health care services extensively, increasingly and in a multidisciplinary manner in the following year.  The model receives input data consisting of an individual's background information and records of their usage of social welfare and healthcare services from the preceding three years. Two machine learning methods were tested in the modelling: logistic regression and gradient-boosted decision tree (XGBoost algorithm). 

The model's classification performance (AUC value) using either of the two machine learning methods was 0.6 on average. The sensitivity of the models varied between 30-40 per cent within the review groups when the specificity was fixed to 75 per cent. Although classification performance and sensitivity remain relatively low, it is important to understand that when we use the prediction model to target anticipatory services, even low classification accuracy and sensitivity can be acceptable and significantly improve the current situation. By allocating anticipatory services based on the prediction model, persons at risk can be reached better than by random selection.

The MAITE project achieved its important goal of producing an overall view of the development of prediction models for use in social welfare and health care services. The demonstration carried out in the project made it possible to illustrate the model to relevant professionals. It provided valuable feedback for further development.

Prediction model suitable particularly for service guidance of elderly people

Based on the feedback, the prediction model could be particularly useful in the service guidance and screening of the elderly for proactive interventions, provided that the application complies with legislation and is integrated into the healthcare provider's existing application environment. 

The project underscored that developing data-driven solutions based on artificial intelligence requires extensive co-operation between different actors and, in this case, the commitment of the user organisation, the Päijät-Häme wellbeing services county.

Using an AI-based model in the operative services also involves legislative perspectives, which have been examined in the MAITE project by the University of Lapland.

The final report of the MAITE project (in Finnish) can be found here.

Authors of this blog:

Principal Scientist Jaakko Lähteenmäki, Senior Scientist Juha Pajula, Research Scientist Heba Sourkatti from VTT