2020-09-14: International Conference on Artificial Intelligence in Medicine (AIME) 2020 Report

2020 International Conference on Artificial Intelligence in Medicine (AIME 2020), hosted by the University of Minnesota, was held virtually 25-28 August 2020.  The first two days were dedicated to tutorials, workshops and the doctoral consortium.  The main conference was held during the last two days. The conference focused on significant theoretical, methodological, and applied results related to the application of artificial intelligence (AI) in medicine.  Although it was held virtually, discussion boards attached to each presentation permitted interactions and networking opportunities with authors.

Day 1 Tutorials

Three tutorials were offered on the first day of the conference:  

  1. Methods and Applications of Natural Language Processing in Medicine
  2. Large Scale Ensembled NLP Systems with Docker and Kubernetes
  3. The Overview Effect: Clinical Medicine and Healthcare Concepts for the Data Scientist.
The first tutorial was held in the morning and the other two were held simultaneously in the afternoon. The AIME 2020 student scholarship that I received included access to one tutorial.  I chose to attend "The Overview Effect: Clinical Medicine and Healthcare Concepts for the Data Scientist" tutorial, which was led by Dr. Anthony C. Chang of the Children’s Hospital of Orange County, California, USA. During the tutorial, Dr. Chang covered various areas related to data science in medicine including, but not limited to, how clinicians think, biases and heuristics common among clinicians, variations in false positives versus false negative acceptances, complexities in biomedicine, and time continuum of disease.  The tutorial made for a very informative afternoon on the first day of the conference.

Day 2 Workshop and Doctoral Consortium

Workshop: AI and the Coronavirus

The morning of day 2 began with the "AI and the Coronavirus" workshop led by the conference co-hosts Dr. Robert Moskovitch from Ben-Gurion University, Israel, and Dr. Martin Michalowski from the University of Minnesota, USA. During his opening remarks, Dr. Moskovitch mentioned that the Journal of Artificial Intelligence Research (JAIR) will launch a special track on AI and the Coronavirus, which will be accepting papers through the end of November 2020. 

Opening remarks were followed by eight presentations relating to the workshop's theme.  Hagit Grushka-Cohen from Ben-Gurion University, Israel, began the presentations with "A Framework for Optimizing COVID-19 Testing Policy Using a Multi Armed Bandit Approach." It was followed by a presentation on "The Impact of Ethnicity, Socioeconomic Status, Air Quality, and Age on COVID-19 Morbidity and Mortality Across USA Counties," which was given by Nevo Itzhak, who was also from Ben-Gurion University, Israel. From the morbidity and mortality rate models that his research group trained, they found that wealthier counties tend to have less COVID-19 cases and related deaths.  However, there was no correlation found between an increase in COVID-19 related deaths and factors like poor air quality.

Following a break halfway through the presentations, the remaining speakers presented their work. Dr. David Buckeridge, a professor in the School of Population of Global Health at McGill University in Canada, presented their "Automating the Analysis of Online Media to Track Public Health Measures against COVID-19" paper.  In this work, dynamic topic modeling was used to analyze time-varying data related to non-pharmaceutical interventions for COVID-19 prevention as obtained from the Global Public Health Intelligence Network (GPHIN) and World Health Organization (WHO). Dr. Maxim Topaz from Columbia University, New York, then briefed his the paper titled "Symptoms Associated with COVID-19:  A Natural Language Processing Study," which discussed findings from a study conducted from two hospitals in New York during the Spring of 2020.  He described his group's method for labeling clinical notes to predict a positive COVID-19 diagnosis. 

Doctoral Consortium

In the afternoon, six PhD students from various universities presented their current and future research during the doctoral consortium, which was held via a BlueJeans video conference.  Mary Borland led the session while Dr. Graciela Gonzalez, Dr. Abeed Sarker, and Dr. John H. Holmes served as mentors.  I was first to present my work on automated Post-Traumatic Disorder Assessment (PTSD) symptom assessment. 

Kathleen Miley then presented her dissertation proposal on "A Data-Driven Approach to Identify the Brain and Cognitive Mechanisms of Social Functioning in Health and Early Schizophrenia."  "A Computational Approaches to Support Automated Conversational Health Coaching" was presented by Elliot Mitchell.  Zhipeng Huang discussed his current and future work on "A Latent Space Model for HLA Compatibility in Kidney Transplantation."  Libby Ferland presented "Enhancing User Modeling in Conversational Agents for Improved Personalization in Elder Support Applications."  ODU's own Gavindya Jayawardena finished the Doctoral Consortium with "RAEMAP: Real-Time Advanced Eye Movements Analysis Pipeline."  

The mentors provided each presenter with a lot of good feedback via email and during the doctoral consortium.

Day 3 (Main Conference Day 1)

The main conference was held during the next two days, which were composed of keynote speakers, panel discussions, paper presentations, poster presentations and awards.  Conference co-chair Dr. Robert Moskovitch started the main conference off with a few opening remarks. 

Keynote Speaker 1 

The first keynote speaker was Dr. Edward H. ("Ted") Shortliffe, who is Chair Emeritus and Adjunct Professor of Biomedical Informatics at Columbia University's College of Physicians and Surgeons in addition to other positions held as noted on his very extensive resume.   His presentation titled "AI Today: Forgetting Our Roots" covered how AI started and how AI in medicine has evolved over his 50 year career.  Dr. Shortlife began with recounting the first use of the term "artificial intelligence" at a conference in 1956.  He identified key AI leaders of the 1960s and early AI efforts, like the Dendral Project, that helped bring about the emergence of AI applications in medicine.

Dr. Shortlife covered significant projects in the history of AI, including the Internist-1, CASNET, and MYCIN projects developed in the 1970s.  The MYCIN project is actually one that Dr. Shortlife worked on during his doctoral dissertation work at Stanford University and his dissertation was later published as a book.  He did note a decline in AI research during the early 1980s before it rebounded later in that decade, but only to decline again the early 1990s.  

Dr. Shortlife discussed how since the late 1990s, AI in medicine has seen continued growth as a result of the integration of decision support functionality in commercial products and a dramatic increase in electronic health records. Together, they would lead to the development of clinical decision support systems that utilize machine learning techniques for data analytics.  These clinical decision support systems provide capabilities that include medical device data interpretation, event monitoring and alerts, and direct consultation with clinical users.

Before concluding his talk, Dr. Shortlife emphasized the need for biomedical informatics and AI in medicine to be treated as a science.  He also believes that industry needs to develop a closer relationship with these areas, which are mergers of medicine and computer science, as it has done with computer science alone. 

Predictive Modeling I Session

The first session of the day included the presentations two papers.  The first paper titled "Lung Cancer Survival Prediction Using Instance-specific Bayesian Networks" was presented by University of Pittsburgh PhD student Fattaneh Jabbari.  Instead of using a population-wide model, she proposed the use of an instance-specific Bayesian Network model for predicting lung cancer outcomes.

Paolo Fraccaro from the Hartree Centre, IBM Research Europe, UK presented their work in support of home monitoring of elderly patient as described in a paper titled "Development and Preliminary Evaluation of a Method for Passive, Privacy-aware Home Care Monitoring Based on 2D LiDAR Data."  Their recommended the use of passive sensors eliminates a lot of the challenges associated with wearables.

Poster Presentations

There were nine poster presentations during this year's conference, which covered a range of topics relating to AI in medicine.  They included Pietro Bosni's poster presentation on "Deep Learning Applied to Blood Glucose Prediction from Flash Glucose Monitoring and Fitbit Data," Olgierd Unold's review of "Unsupervised Grammar Induction for Revealing the Internal Structure of Protein Sequence Motifs," and Taylor-James Gilard's brief of his "AI Medical School Tutor: Modelling and Implementation" poster.

Presenters were only given five minutes for their presentations.  However, conference attendees were encouraged to further engage poster presenters and ask questions in the virtual poster session lobby during the lunch break scheduled that afternoon.

Unsupervised Learning and Bioinformatics Session

A couple authors presented during the unsupervised learning and bioinformatics session.  One of the two was Matthew Casey from Ardsley High School, USA.  He presented his work on the "Analysis of Viability of TCGA and GTEx Gene Expression for Gleason Grade Identification."  Mr. Casey used a RNA sequencing data classification method to improve the accuracy of the Gleason Grade score identified for patients in support of prostate cancer diagnosis.

Temporal Data Analysis I Session

The first Temporal Data Analysis session began with a presentation from Jeong Min Lee from the University of Pittsburgh, USA, on "Multi-scale Temporal Memory for Clinical Event Time-Series Prediction."  The objective of his work is to develop a model that could predict a patient's future condition within a certain time window to allow physicians to better respond to patients in their care.

Mr. Lee's presentation was followed by the presentation of a paper that would later win the best student paper award titled "Hype:  Predicting Blood Pressure from Photoplethysmograms in Hypertensive Population." The paper was briefed by Ariane Morassi Sasso from the Digital Health Center, Hasso Plattner Institute, University of Potsdam.  A little later, Tianran Zhang from UCLA described their work on "Diagnostic Prediction with Sequence-of-sets Representation Learning for Clinical Event."

Predictive Modeling II Session

The second Predictive Modeling session of the day included four paper presentations, two of which were best paper candidates.  Roshan Tourani, from the Institute for Health Informatics, University of Minnesota, USA, presented their best paper candidate, "Consensus Modeling: A Transfer Learning Approach for Small Health Systems."

Dr. Tomer Sagi from the Department of Computer Science at Aalborg University, Denmark, later presented his group's best paper candidate titled "Towards Assigning Diagnosis Codes using Medication History."  During the presentation, he described the architecture of the system developed by his group to predict patients' conditions based on their medication history.

Clinical Practice Guidelines Session

During the Clinical Practice Guidelines session of the conference, Nick Fung from University of Twente, The Netherlands presented their work on "A Verified, Executable Formalism for Resilient and Pervasive Guideline-based Decision Support for Patients," which utilizes a data flow model for disease management from their previous work known as MADE. Dr. William Van Woensel,Computer Faculty member at Dalhousie University, Canada, discussed their work on "A CIG Integration Framework to Provide Decision Support for Comorbid Conditions using Transaction-based Semantics and Temporal Planning."

Deep Learning Session

The conference continued into the evening with the last session of the day on Deep Learning. One of the presenters, Tingyi Wanyan from the Icahn School of Medicine at Mount Sinai, who also represented Indiana University and the University of Texas Austin, shared his work as detailed in the paper "Heterogenous Graph Embeddings of Electronic Health Records Improve Critical Care Disease Predictions."

Gabriel Schamberg also presented the paper "Controlling Level of Unconsciousness by Titrating Propofol with Deep Reinforcement Learning." During his presentation, Mr. Schamberg described a system for managing unconsciousness associated with general anesthesia. The objective is to have a sensor, such as EEG, determine when another dose of propofol is administered to a patient to maintain a target level of unconsciousness.  Additional work and simulation studies may one day lead to a live system.

One more presenter would share his work before the conclusion of the first day of the main conference.

Day 4 (Main Conference Day 2)

Panel Discussion

The final day of the conference began with a panel discussion on Artificial Intelligence in the Time of Coronavirus.  The panel included Dr. David Buckeridge of McGill University, Canada; Dr. Mengchun Gong from Digital China Health Technology, China; Dr. Laurence Lovat from the University College London Hospitals, UK; Dr. Orly Weinstein with Clalit Healthcare Services, Israel; and Dr. Glenn Kramer of xCures, USA.

Keynote Speaker 2

Following the panel discussion, Dr. Moskovitch introduced the second keynote speaker of the conference, Dr. Vimla L. Patel who is a Senior Research Scientist and Director at the Center for Cognitive Studies in Medicine and Public Health at the New York Academy of Medicine.  She is also an adjunct professor at three different universities:  Columbia University, Arizona State University, and Weill Cornell College of Medicine.  Dr. Patel was invited to give her keynote speech on "Human Cognition: A Guide to Evolution of AI."

Dr. Patel discussed how cognitive sciences including philosophy, linguistics, cognitive psychology, neurosciences, and computer science could be used to better understand the brain and its cognitive processes such as logic, problem solving, reasoning, perception, intelligence, creativity, memory, learning, and language.  She proposed looking for synergies between AI and cognitive psychology as well as using techniques in one area to influence the other and evolve ideas and solutions.

Dr. Patel identified those people and studies that had influenced her work. She also described her efforts in support of creating a bridge between cognition and medical AI including her clinical work, publications and speaking engagements.  Dr. Patel gave her thoughts on the future role of cognition in AI.  She also emphasized the need for more collaboration between AI and medicine researchers before eventually taking questions from the audience.

Temporal Data Analysis II Session

Four authors presented their works during the second Temporal Data Analysis session of the conference.  That included a presentation by Nevo Itzhak from Ben-Gurion University of the Negev, Israel, on "Acute Hypertensive Episodes Prediction," a best student paper candidate. He and his group determined the performance of multiple prediction models when used with vital sign data to predict the pre-hypertensive stage of acute hypertensive episodes. 

Other presenters during this session discussed their methods for falls prediction in self care homes, medical time-series data generation, and predicting time-to-event outcomes. 

Natural Language Processing Session

Lunch was followed by a two-hour session on Natural Language Processing.  It began with a presentation of the paper, "Comparing NLP Systems to Extract Entities of Eligibility Criteria in Dietary Supplements Clinical Trials" by one of its authors, Anusha Bompelli from the Institute for Health Informatics, University of Minnesota, USA.  

Dmitry Umerenkov from the Sberbank Artificial Intelligence Laboratory, Russia, presented his work titled "Predicting Clinical Diagnosis from Patients Electronic Health Records Using BERT-based Neural Networks." A few presentations later, PhD Student Yves Mercadier from Université de Montpellier, France, presented one of the best student paper candidates, "Divide to Better Classify."  Since 80% of health care information is stored in text format but often in small databases, his group sought to improve the learning of neural networks for these smaller databases through data augmentation.

Information Retrieval and Image Processing Session

The last session of this year's conference included five presentations that centered around information retrieval and image processing.  The first of those presenters was Dr. Hamid Tizhoosh from the University of Waterloo, Canada, who detailed a process that uses deep neural networks for image feature extraction and classification to detect pneumothorax, or a collapsed lung, which is often hard to detect by an untrained eye.

Dr. Tizhoosh also presented two other papers: "A New Local Radon-Based Descriptors for Content-Based Image Search" and "Forming Local Intersection of Projections for Classifying and Searching Histopathology Images."

Jerry Wei from Darmouth College, USA, reviewed his group's paper, "Difficulty Translation in Histopathology Images."  He described a method for modifying histopathology images to make them harder to classify then using those augmented images to improve classifier performance. Finally, work on "Weakly-Supervised Segmentation for Disease Localization in Chest X-ray Images" was presented by Ostap Viniavskyi and Mariia Dobko from the Machine Learning Lab at Ukrainian Catholic University


Co-host Dr. Martin Michalowski closed out this year's conference with a quick recap of the conference events and presentation of the best paper awards.  The AIME 2020 Best Paper winner was "Controlling Level of Unconsciousness by Titrating Propofol with Deep Reinforcement Learning" by Gabriel Schamberg, Marcus Badgeley and Emery N. Brown.  The AIME 2020 Best Student Paper winner was titled "Hype:  Predicting Blood Pressure from Photoplethysmograms in Hypertensive Population," written by Ariane Morassi Sasso, Suparno Datta, Michael Jeitler, Christian S. Kessler, Bert Arnrich, and Erwin Boettinger.  Best papers will be invited to submit extended manuscripts for inclusion in a special issue of Artificial Intelligence in Medicine (AIIM).  Dr. Michalowski also announced the dates and location of next year's AIME conference.

Although the conference has come to an end, conference organizers did make recordings of the conference available to attendees for several months.  With that, we look forward to AIME 2021

-- Bathsheba Farrow