AI for Patient Diagnosis and Treatment

Oncology Patients

In collaboration with Roche Diagnostics GmbH

  • Roche Diagnostic is a Fortune-100 company focused on medical diagnostic applications.
  • Our aim is to design a super-learner by combining deep learning, survival analysis, and random forest for best longitudinal estimation of survival time. We remove confounders and design causally-informed model selection processes.
  • FlatIron database, acquired by Roche, is used for this project and contains EHR of 2.4M oncology patients.

COVID-19 Patients

In collaboration with the department of Internal Medicine-Pulmonology at the Ludwig Maximilian University Hospital (Germany) and the department of Emergency Medicine at the Johns Hopkins Hospital (USA)

  • Our aim is to develop and validate AI models that can reliably predict the course of COVID- 19 progression and its longer-term (6-months) complications via modeling patients' clinical trajectories to understand their recovery and development of pulmonary complications.
  • We integrate both clinical and biological signatures (via virome, mirobiome and genome analysis) to build a full picture of the disease.
  • In addition, we compare trajectories of COVID-19 patients to those recovering from similar- symptom diseases such as influenza and community-acquired pneumonia.
  • We further develop a decision support system for physicians to allocate resources and to plan individualized follow-up care for at-risk patients.
  • For this project, we collect clinical data and bio-samples (admission to 6 months post- discharge) from five different international patient cohorts (three in Munich and two in the USA).

Cardiac Surgery Patients

In collaboration with the Cardiac Surgery at the Johns Hopkins Hospital and Malone Center for Engineering in Healthcare

  • Our first aim is to develop AI models that identify with high precision and recall, and days in advance, patients on trajectories toward developing adverse outcomes.
  • Our second aim is to disentangle the models from confounders, such as treatment/medication decisions, via causal-inference analysis and to measure the actual impact of decisions on the adverse outcomes.
  • HIT (heparin-induced thrombocytopenia, a form of blood clotting due to prescribed medication) is an example of partially-observed adverse outcomes. Onset of HIT, due to late or missed diagnosis, is a potentially catastrophic clinical event, leading to brain stroke, organ failure, limb amputation, and death.
  • The aim of the HIT project is to discover mis-diagnosed patients via developing a machine- learning model (semi-supervised) to identify patients with a probable diagnosis of HIT who failed to have a standard HIT-workup during their hospital stay, meaning that their severe condition was not diagnosed during their stay.
  • Early prediction of ICU bounce-back (return to ICU after misdiagnosis of recovery stage), acute kidney failure, prolonged intubation, and readmission are other ongoing similar projects with fully-observed adverse outcomes.
  • Comprehensive EHR data from 8,000 cardiac surgery patients (adults and pediatric) are available.

Preterm Neonate Patients

In collaboration with the Neonate Department at the Ludwig Maximilian University Hospital

  • Our aim is to design models for early prediction of short- and long-term neurological, cardiovascular, and pulmonary co-morbidities in preterm infants, via analysis of their combined electronic health records, omics, and imaging data.
  • In addition, we aim to measure treatments effectiveness via counterfactual analysis to measure whether a proposed change of treatment would have prevented the undesired outcome.
  • Data is collected over eight years from 700 preterm infants in the form of comprehensive EHR (during the NICU stay plus five years of follow up), Omics-based features (bio-samples taken from mother, infant, and environment used for proteome, virome, microbiome analysis), and imaging data (brain/lung/heart MRI scans).

Chronic Lung Disease Patients

In collaboration with the Ludwig Maximilian University Hospital and the German Center for Lung Research

  • Chronic lung diseases are the 4th-leading cause of death worldwide.
  • The health conditions even of monitored patients can deteriorate acutely and unforeseeably, leading to severe complications (aka. exacerbation).
  • Our aim is to develop AI models to identify risk indicators of exacerbation, define early predictors, and measure whether early detection of exacerbation is associated with improved outcomes.
  • Deeply-phenotyped and EHR data of 8,000 CLD patients (tracked since 2013) is provided by Comprehensive Pneumology Center in Munich.

Stay Tuned for Other Projects ...