Bringing machine learning to the clinic

The primary goal of our lab is to develop computational analyses and tools that reach patient care to improve  outcomes. The convergence of machine learning and expansive routine clinical data enable the development of automated predictive models which can be integrated into the clinical workflow to personalize care.

We have developed predictive tools with a range of complexity, including identifying favorable characteristics in patients with oligometastatic cancers (Hong, PLOS ONE 2018) and predicting testosterone recovery following androgen deprivation (Spiegel, Int J Radiat Oncol Biol Phys 2019). Among these, we developed an electronic health record-based machine learning algorithm to predict acute care visits during outpatient cancer therapy (Hong, JCO Clin Cancer Inform 2018). This tool was recently evaluated in one of the first prospective randomized studies of clinical machine learning implementation (results are pending).

Learning from real world data

Real world data offers insights into practice patterns and may generate knowledge where prospective trials are not feasible or do not sufficiently generalize to patients. We analyze data from electronic health records, national databases, institutional experiences, claims data, and payor data to both learn from prior experiences and characterize data limitations.

We have evaluated practice patterns in a range of cancers including lung (Hong, J Natl Compr Canc Netw 2019), prostate, and gynecologic malignancies. Among these studies, we have used Medicare provider data to identify disparities in access to specialized therapies for prostate and gynecologic malignancies (Hong, Brachytherapy 2018).

Synthesizing multi-modal data

Routine clinical practice generates multi-modal complementary data, including imaging, pathology slides, laboratory time series, and -omics. We apply methods to generate quantitative features from this complementary data to generate new analyses.

Imaging can be converted into quantitative features which can complement other sources of data. Additionally, it may offer a proxy for other forms of clinical data. We have applied quantitative imaging approaches to a number of oncologic questions, including prediction of outcomes by PET/CT during treatment of anal (Hong, Int J Radiat Oncol Biol Phys 2018) and esophageal (Tandberg, Int J Radiat Oncol Biol Phys 2018) cancers, as well as other clinical applications such as the approximation of pulmonary function tests based on CT (Lafata, Sci Rep 2019).

Unlocking unstructured clinical data

Large amounts of data are captured in free-text clinical notes, which can pose challenges to clinical study. Extracting this unstructured information can both inform clinical research and facilitate deployable computational tools.

We have adapted clinical natural language processing pipelines to extract patient outcomes during cancer therapy from routine clinical documentation. Our studies indicate high variability even among human experts, which may motivate the use of systematic computational approaches (Fairchild, Int J Radiat Oncol Biol Phys 2020).