Predicting comorbidities of epilepsy patients using big data from Electronic Health Records combined with biomedical knowledge

Abstract

Epilepsy is a complex brain disorder characterized by repetitive seizure events. Epilepsy patients often suffer from various and severe physical and psychological comorbidities. While general comorbidity prevalence and incidences can be estimated from epidemiological data, such an approach does not take into account that actual patient specific risks can depend on various individual factors, including medication. This motivates to develop a machine learning approach for predicting individual comorbidities. To address these needs we used Big Data from electronic health records (˜100 Million raw observations), which provide a time resolved view on an individual's disease and medication history. A specific contribution of this work is an integration of these data with information from 14 biomedical sources (DisGeNET, TTD, KEGG, Wiki Pathways, DrugBank, SIDER, Gene Ontology, Human Protein Atlas, ...) to capture putative biological effects of observed diseases and applied medications. In consequence we extracted >165,000 features describing the longitudinal patient journey of >10,000 adult epilepsy patients. We used maximum-relevance-minimum-redundancy feature selection in combination with Random Survival Forests (RSF) for predicting the risk of 9 major comorbidities after first epilepsy diagnosis with high cross-validated C-indices of 76 -89% and analyzed the influence of medications on the risk to develop specific comorbidities. Altogether we see our work as a first step towards earlier detection and better prevention of common comorbidities of epilepsy patients.

An Explainable Multi-Modal Neural Network Architecture for Predicting Epilepsy Comorbidities Based on Administrative Claims Data

“…We start by explaining the principle architecture of DeepLORI. In agreement to our former work, one of the key ideas is that claims data has an inherent hierarchical structure (Gerlach, Lu and Fröhlich, 2017): The data initially contains three major types of features: 1) prescribed substance codes, 2) diagnoses codes (mapped to PheWAS terms, see above) and 3) general demographic information, such as age, gender and major metropolitan area information. Monthly reported prescriptions and diagnoses can typically be represented via a one-hot vector encoding.…”

“…We compared DeepLORI against several competing approaches: 1. Random Survival Forests (Ishwaran et al, 2008): In this earlier published approach (Gerlach, Lu and Fröhlich, 2017) we first combined claims data with biomedical knowledge (akin to this paper) and then used a window of fixed length (3 months) to summarize features via a max-pooling. Features encoding prescriptions and diagnoses within such a time window were concatenated, resulting into an overall number of around 165,000 features per patient.…”

“…We used two cohorts, (1) the original data covering years 2011-2015 for model-training and –evaluation within a nested cross-validation scheme, and (2) the external validation data covering years 2008-2018 to validate the models trained on the “original data”. In agreement to our earlier publication (Gerlach, Lu and Fröhlich, 2017) and common practice at UCB, epilepsy patients in the original data were identified matching at least one of the following criteria:…”

“…For epilepsy, (Glauser et al, 2020) proposed an ML model for psychatric comorbidities based on survey data from 122 patients. In our earlier work (Gerlach, Lu and Fröhlich, 2017) we proposed an ML model (Random Survival Forests) using US administrative health claims data from ~10,000 epilepsy patients to predict several major comorbidities (anxiety, bipolar disorder & schizophrenia, diabetes type 2, migraine, overweight & obesity, stroke & ischemic attacks) of epilepsy patients. Administrative health claims data have generally been shown useful for developing ML models in the epilepsy field.…”

“…The opportunities of healthcare claims data for ML based modeling have further been discussed in (Fröhlich et al, 2018;Miotto et al, 2018;Xiao, Choi and Sun, 2018;Thesmar et al, 2019;Kwak and Hui, 2020). In our earlier work we demonstrated the possibility to augment claims data with biomedical background knowledge, hence enabling the interpretation of machine learning models down to the level of disease associated biological processes (Gerlach, Lu and Fröhlich, 2017). The particular novelty of the present work is a dedicated multi-modal neural network architecture for administrative claims data, which we call Deep personalized LOngitudinal convolutional RIsk model (DeepLORI).…”