This talk will present PreFree - a machine learning powered decision support and home monitoring tool that helps to prevent severe complications in pregnancy. PreFree aims to assist physicians in deciding whether to send a patient home with monitoring or to hospitalize her. Focus of this talk is the training and application of our ML models based on clinical outcome data from the Charite Berlin.
PreFree is a decision support and home monitoring tool, consisting of machine learning models to predict severe complications in pregnancies and mobile applications to monitor vitalsigns of pregnant women staying at home.
We will go into detail on how we trained our models and how they are put into use. We will talk about the most interesting challenges along the way of training them: Dealing with dirty, small and wide data, generalization, interpretation and putting them into production.
Medical use case
With an incidence of 2% to 5%, preeclampsia is the leading cause of maternal morbidity and mortality worldwide. There is an unmet need to early detect patients at-risk especially for preeclampsia-related severe outcomes.
PreFree aims to do this with machine learning (ML) models that rely on a multitude of inputs, from special bloodmarkers over vital signs to diameters of arterias measured via ultra sound.
When women with signs and symptoms of preeclamspia present at a clinic, physicians face a problem: They have to decide based on a very inprecise gold-standard, wether they should hospitalize the women with signs and symptoms of preclampsia or send them home.
Our ML models aim to support the decision of the physician through predicting wether a severe outcome from the field of preclampsia is likely to occur or not.
Our training data contains ~2500 measurements from ~1600 women with signs and symptoms who presented at the Charité in Berlin from 2010 to 2019. For each woman, all existing measurements and outcome data was manually collected into an excel sheet which ended up having a total of ~250 columns. Therefore some of our first challenges where to identify the most predictive features and clean the data as well as possible. We will go into detail on how we did that and describe our further way to building and validating our models that we now plan to evaluate in clinical trials. We will furthermore talk about our efforts to improve generalization. Finally, we will describe how we are putting them into production.
This talk will be of informative and light-hearted character, even though knowledge about machine learning methods might be beneficial. Our main goal is to share our experiences on building a machine learning powered tool in health tech.