Health Data Science
Mining clinical charts for immune-related adverse events
Immunotherapy is a type of cancer treatment that boosts the body’s natural defenses to fight cancer. Immune checkpoint inhibitors (ICIs) are a type of immunotherapy drugs that work by blocking immune checkpoints (the “brakes” of the immune system) that cancer cells frequently manipulate to protect themselves. Thus, ICIs can stimulate new immune responses to promote the elimination of cancer cells and can lead to substantially improved survival in patients with advanced malignancies. However, the manipulation of the immune system’s brakes via ICIs may lead to overactive immune responses, which may lead to side effects such as rash, diarrhea, fatigue, colitis pneumonitis, etc. These side effects due to immunotherapy drugs are termed as Immune-Related Adverse Events (irAEs). These adverse events can sometimes be severe and may diminish the benefits of immunotherapy. To ensure treatment safety, research efforts are needed to comprehensively detect and understand irAEs.
Knowledge about adverse events in general and irAEs specifically is mostly buried in clinical charts in electronic databases. Methodologies that could automate the recording of irAEs from clinical charts would benefit accurate cohort identification for clinical trials and accelerate the creation of curated datasets for developing predictive models. To address such needs, we have developed a Natural Language Processing pipeline to identify patients treated ICIs as having irAEs or not based on their longitudinal set of clinical notes. We specifically compare the performance of different machine learning models (shallow and deep learning models) using different feature representations (frequency-based, distributed word embeddings) and text reduction (keyword-based filtering) methods. Our work demonstrates that deep learning models with text-reduction using keyword filtering and word embeddings can achieve good accuracy in classifying patients with irAEs based on the clinical notes.
Read more about this work in the medrxiv preprint server for Health Sciences
Opioid overuse prediction
Patients who undergo upper extremity surgeries are at substantial risk of developing opioid use disorder after prescribed postoperative use. While early interventions have shown promise in reducing opioid consumption, many patients still have high opioid use after surgery.
In collaboration with the Curtis National Hand Center, in Baltimore Maryland, data scientists at ICBI are investigating the association between patient reported data and postoperative pain control challenges in over 2000 patients who underwent surgery at MedStar Health System. We are using state of the art machine learning techniques to predict opioid use status at clinically relevant time points throughout a patient’s postoperative period. The predictive models leverage a combination of patient reported data and electronic health records.
Claims data page
Information derived from health claims and insurance plans data can provide insights into a patient and a population health status and inform healthcare plans and providers on patient health trajectories, potential interventions and cost efficiency. Claims data include patient demographics, procedures, medical conditions, prescription drugs and healthcare resources used for specific conditions over extended periods of time. Read more here (new window).