Deep Learning in Healthcare Summit schedule

Natural Language Processing for Healthcare

With recent advancements in Deep Learning followed by successful deployment in natural language processing (NLP) applications such as language understanding, modeling, and translation, the general hope was to achieve yet another success in healthcare domain. Given the vast amount of healthcare data captured in Electronic Medical Records (EMR) in an unstructured fashion, there is an immediate high demand for NLP to facilitate automatic extraction and structuring of clinical data for decision support. Nevertheless, the performance of off-the-shelf NLP on healthcare data has been disappointing. Recently, tremendous efforts have been dedicated by NLP research pioneers to adapt general language NLP for healthcare domain. This talk aims to review current challenges researchers face, and furthermore, reviews some of the most recent success stories.

3 Key Takeaways:

*General overview of state-of-the-art NLP

*How to build a domain-specific NLP pipeline for life science applications

*Review of a few successful applications of NLP in life sciences and how the future will/should look

Amir Tahmasebi is the director of Deep Learning at Enlitic, San Francisco, CA. Before joining Enlitic, Amir was the senior director of Machine Learning and AI at CODAMETRIX, Boston, MA. He also served as a lecturer at MIT, Northeastern University, Boston University, and Columbia University. Prior to CODAMETRIX, Dr. Tahmasebi was a Principal Research Engineer at PHILIPS HealthTech, Cambridge, MA. Dr. Tahmasebi’s research is focused on innovating computer vision and natural language processing solutions for patient clinical context extraction and modeling, clinical outcome analytics and clinical decision support. Dr. Tahmasebi received his PhD degree in Computer Science from the School of Computing, Queen's University, Canada. He is the recipient of the IEEE Best PhD Thesis award and Tanenbaum Post-doctoral Research Fellowship award. He has been serving as area chair for MICCAI and IPCAI conferences. Dr. Tahmasebi has published and presented his work in a number of conferences and journals including NeurIPS, NAACL, MICCAI, IPCAI, IEEE TMI, SPIE, and RSNA. He has also been granted more than 15 patent awards.

Using AI-Guided Analytics in Early Stage Clinical Trials

The Pharmaceutical and Health IT fields have struggled to analyze the high volume and diverse types of molecular and clinical data currently available. In this regard, AI informed discovery presents a potential solution over the more traditional expert guided approaches. By favoring interpretability, continued method optimization and data flexibility, Berg Analytics has developed a data-driven and hypothesis-free AI solution that has been successfully utilized to generate actionable outcomes in both clinical development and healthcare fields. In an effort to take advantage of both probabilistic and neural net-based AI technologies, Berg Analytics is currently exploring methods to augment these deep-learning methods.

Dr. Eric Milliman, Ph.D. is a Data Scientist at Berg Analytics, a division of Berg Pharma, where he works on researching, developing and applying innovative data-driven methods in the de-novo discovery of therapeutics and biomarkers. Prior to his role at Berg, Dr. Milliman was a post-doctoral fellow in the Epigenetics & Stem Cell Biology Laboratory at the National Institute of Environmental Health Sciences. Dr. Milliman holds a Ph.D. in Biology from the State University of New York at Buffalo.

Accelerating High Throughput Drug Discovery Using Deep Learning

Image-based high content screening is a well-established method for discovering new compounds at high throughput using their phenotypic signatures. Current processing of these imaging assays relies primarily on tedious and often subjective manual feature extraction requiring specialized skills. We've explored the application of deep convolutional neural networks to address these concerns. Beyond algorithmic hurdles, we faced challenges around building robust production-level models as well as integrating within a pharmaceutical context. This talk will discuss our approach towards building and industrializing deep learning based image analysis workflows within the early phases of drug discovery research.

Yusuf works as a data scientist in drug discovery research at GlaxoSmithKline in Cambridge, MA. Currently, his main focus is on putting together a computer vision platform for early stage drug discovery with broad usability across use cases and imaging domains. He also has previous experience within the areas of compound screening and biomarker identification as well as in building systems that fit a healthcare context. Yusuf was previously employed at Merrimack Pharmaceuticals and received his M.S. from Carnegie Mellon in 2015.

DeepVariant: Highly Accurate Genomes With Deep Neural Networks

Deep learning has enabled dramatic advances in image recognition performance. In this talk I will discuss using a deep convolutional neural network to detect genetic variation in aligned next-generation sequencing human read data. Our method, called DeepVariant, both outperforms existing genotyping tools and generalizes across genome builds, sample preparations, and sequencing instruments. DeepVariant represents a significant step from expert-driven statistical modeling towards more automatic deep learning approaches for developing software to interpret biological instrumentation data.

Cory McLean is a Senior Software Engineer on the genomics team at Google Brain. His research interests include the development and application of machine learning methods to genomics and the biology of human health. Recent projects include efforts to use deep learning image classification methods to improve the detection of genetic variation from high-throughput sequencing data and to interpret cellular morphology and function. Before joining Google, Cory was a research scientist at 23andMe where he studied Parkinson's disease and population genetics. He received his B.S. and M.Eng. in computer science from MIT and his M.S. and Ph.D. in computer science from Stanford University.

Prediction of Opioid Cessation Using Machine Learning

The non-prescribed use of opioid analgesics has become a significant health and social problem. While studies had focused on opioid use disorder (OUD), a systematic inquiry of the lifestyle factors for opioid cessation has not been conducted. We performed supervised machine learning with feature selection on Yale-Penn dataset among subjects who met DSM-5 OUD criteria to predict current opioid use status. Subjects were interviewed by Semi-Structured Assessment for Drug Dependence and Alcoholism (SSADDA), a comprehensive diagnostic questionnaire. We observed moderate to high prediction efficiency using LASSO, SVM and random forest with SSADDA input alone. Future steps include using DNN and adding genetics from GWAS to see the improvement of accuracy.

Jiayi Wu Cox is a Ph. D candidate in Genetics and Genomics program at Boston University. Her research focuses on using bioinformatics and machine learning to find the genetic risk factors for a variety of diseases including addiction. In addition to her thesis, she also studies epigenetic changes in response to drug treatment and gene expression level differences due to disease phenotypes. Jiayi graduated from Tufts University with M.S degree in Pharmacology and Experimental Therapeutics. She hope her work can facilitate the development of personalized medicine, that people will be treated based on who they are genetically and the lifestyle they choose.

Deep Patient: Predict the Medical Future of Patients with Deep Learning and EHRs

The latest advances in deep learning provide new effective opportunities to model, represent, and learn from large amounts of heterogeneous medical data. Here, in this talk, we focus on applying deep learning to the electronic health records (EHRs). In particular, we review the data as well as the recent literature, we highlight limitations and needs for improved methods and applications, and we discuss the challenges to implement and deploy machine intelligence into the clinical domain. We then present Deep Patient, a general-purpose patient and phenotype representation derived from the EHRs that facilitates clinical predictive modeling and medical analysis, such as patient stratification and disease definition.

Riccardo Miotto is a senior data scientist in the Department of Genetics and Genomic Sciences at the Icahn School of Medicine at Mount Sinai in New York and a member of the Institute for Next Generation Healthcare. Riccardo’s work encompasses the design of algorithms for information retrieval and machine learning applied to clinical data for personalized medicine. Previously, Riccardo worked on clinical trial search engines through free-text eligibility criteria processing and machine learning applied to music information retrieval, in the particular semantic discovery and recommendation, automatic tagging, and cover identification. Riccardo obtained his Ph.D. in Information Engineering from the University of Padova, Italy.

Deep Learning-based Clinical Temporal Relation Extraction

The extraction of temporal relations in medical text has been drawing growing attention because of its potential to dramatically increase the understanding of many medical phenomena such as disease progression, longitudinal effects of medications, a patient's clinical course, and its many clinical applications such as question answering, clinical outcomes prediction, and the recognition of temporal patterns and timelines. We develop deep neural architectures (e.g. Convolutional or Recurrent Neural Networks) for clinical temporal relation extraction. Comparing with conventional models, which make use of heavily engineered features, neural models employ simple token features or slightly enhanced features, establishing state-of-the-art results.

Chen Lin is an Informatician in Boston Children’s Hospital Informatics Program-Natural Language Processing (CHIP-NLP) group, an Apache cTAKES developer. Chen’s work involves using machine learning in clinical NLP tasks. Topics include clinical temporal relation extraction, disease activity classification based on clinical narratives, drug-induced adverse event prediction, deep phenotyping. His main interests in computational methods include convolutional neural networks, recurrent neural networks, auto-encoder/decoder, word and syntactic embeddings, semi-supervised learning, feature selection, and kernel methods. He holds Master degree in Computer Science from Brandeis University.

AI for Care Planning Support

Effective care planning requires care managers to understand patient health status and needs to deliver appropriate patient support. The proliferation of healthcare data including massive volumes of clinical free text documents, creates a significant challenge for care managers, but a major opportunity for advanced clinical analytics. Novel Artificial Intelligence (AI)-driven solutions can help optimize care planning, reducing inefficiency and increasing focus on the most salient information, leading to improved patient outcomes. This talk will focus on various deep learning-based clinical natural language processing use cases developed as part of our advanced care planning initiatives.

Key Takeaways:

*Effective care planning requires care managers to understand patient health status and needs to deliver appropriate support

*Clinical domain has unique challenges such as massive structured/unstructured data, redundancy, limited interoperability, widespread use of acronyms etc.

*AI-augmented solutions can help optimize care planning, reducing inefficiency and increasing focus on the most salient information leading to improved patient outcomes

Dr. Sadid Hasan is a Senior Director for AI at CVS Health leading the team responsible for AI-enabled clinical care plan initiatives in Aetna. His recent work involves solving problems related to clinical information extraction, paraphrase generation, natural language inference, and clinical question answering using Deep Learning. Sadid has over 60 peer-reviewed publications in the top NLP/Machine Learning venues, where he also regularly serves as a program committee member/area chair including ACL, IJCAI, EMNLP, NeurIPS, ICML, COLING, NAACL, AMIA, MLHC, MEDINFO, ICLR, ClinicalNLP, TKDE, JAIR etc. 

Clinical Motion Lab in Your Pocket

Many neurological and musculoskeletal diseases impair movement, which limits people’s function and social participation. Quantitative assessment of motion is critical to medical decision-making but is currently possible only with expensive motion capture systems and highly trained personnel. We developed AI-based algorithms for quantifying gait pathology using commodity cameras. Our methods increase access to quantitative motion analysis in clinics and at home and enable researchers to conduct studies of neurological and musculoskeletal disorders at an unprecedented scale.

3 Key Takeaways:

*Quantitative assessment of movement enables diagnostics and treatment of many neurological disorders

*Existing methods for quantitative analysis of movement require very expensive equipment

*Deep learning models can predict common gait metrics using a mobile phone camera

Łukasz Kidziński is a co-founder of Saliency and a research associate in the Neuromuscular Biomechanics Lab at Stanford University, applying state-of-the-art computer vision and reinforcement learning algorithms for improving clinical decisions and treatments. Previously he was a researcher in the CHILI group, Computer-Human Interaction in Learning and Instruction, at the EPFL in Switzerland, where he was developing methods for measuring and improving engagement of users in massive online open courses. He obtained a Ph.D. degree at Université Libre de Bruxelles in mathematical statistics, working on frequency-domain methods for dimensionality reduction in time series.

Predicting Individual Physiologically Acceptable States at Discharge from a Pediatric Intensive Care Unit

We aim to quantify acceptable ICU-discharge vitals, compare to age-normal vitals, and develop machine learning models to predict throughout each patient’s ICU episode. Our dataset of 7,256 surviving PICU episodes (5,632 patients) collected between 2009 and 2016 at Children's Hospital Los Angeles contains 375 variables representing vitals, labs, interventions, drugs, and medical and physical discharge times. The means of each patient's heart rate, systolic blood pressure, and diastolic blood pressure between medical and physical discharge were computed as their physiologically acceptable state space (PASS), which were compared to age-normal, regression, and recurrent neural network (RNN) predictions.

Mr. David Ledbetter has an extensive and deep understanding of decision theory. He has experience implementing various decision engines, including convolutional neural networks, random forests, extra trees, and linear discrimination analysis. His particular area of focus is in performance estimation, where he has demonstrated a tremendous ability to accurately predict performance on new data in nonstationary, real-world scenarios. David has worked on a number of real-world detection projects, including detecting circulating tumor cells in blood, automatic target recognition utilizing CNNs from satellite imagery, make/model car classification for the Los Angeles Police Department using CNNs, and acoustic right whale call detection from underwater sonobuoys. Recently, David has been developing an RNN to generate personalized treatment recommendations to optimize patient outcomes using electronic medical records from 15 years of data collected from the Children's Hospital Los Angeles Pediatric Intensive Care Unit.

Image-based Morphological Profiling Using Deep Learning

Microscopy images are widely used in biological research to understand cellular structure. Images can be used to measure multiple properties of cells, and then model the effect that treatments produce at the cellular level. These measurements can be used to discover novel relationships between treatments. We investigate whether relevant cellular features can be learned automatically from images using deep learning. In this talk, we present the results of our experiments, which indicate that deep learning features can improve the ability to identify unknown biological relationships.

Juan Caicedo is a postdoctoral researcher at the Broad Institute of MIT and Harvard, where he investigates the use of deep learning to analyze microscopy images. Previous to this, he studied object detection problems in large scale image collections also using deep learning, at the University of Illinois in Urbana-Champaign. Juan completed research internships in Google Research, Microsoft Research, and Queen Mary University of London in the past, studying problems related to large scale image classification, image enhancement, and medical image analysis. His research interest include computer vision, machine learning and computational biology.

From Pixels to Clinical Insight

Medical images contain a wealth of information about a patient but require a human expert to look at each image to capture relevant clinical knowledge. Computational tools that extract clinically important information from images enable development of novel biomarkers of disease, support surgical planning, and enable disease prognosis. From monitoring fetal development, to predicting stroke outcomes, this talk will discuss current challenges in extracting clinically actionable information from images.

Polina Golland is a professor of EECS at MIT CSAIL. She received her PhD from MIT and her Bachelor and Masters degree from Technion, Israel. Polina's primary research interest is in developing novel techniques for medical image analysis and understanding. With her students, she has demonstrated novel approaches to image segmentation, shape analysis, functional image analysis and population studies. Polina has served as an associate editor of the IEEE Transactions on Medical Imaging and of the IEEE Transactions on Pattern Analysis and Machine Intelligence. She is a Fellow of the International Society for Medical Image Computing and Computer Assisted Interventions.

Deep Learning Opportunities in Cancer Imaging

Radiographic medical images contain a vast wealth of information allowing for accurate non-invasive tumor characterization and ultimately improving cancer diagnosis and care. Recent advances in AI, deep learning in particular, promise to impact multiple facets of the radiology profession and support clinical decision making. The Computational Imaging and Bioinformatics Lab at Dana Farber Cancer Institute and Harvard Medical School is a data science lab focused on the development and application of AI methods on medical data.

In this talk, we will be presenting case studies investigating the clinical utility of deep learning in the detection, segmentation and outcome prediction of cancer tumors. We will also be presenting modelhub.ai, a repository of deep learning models crowdsourced through contributions by the scientific research community.

Ahmed Hosny is a machine learning research scientist focused on solving biomedical problems. Currently at Dana Farber Cancer institute, he trains and optimizes deep learning networks for the prognostication and treatment response prediction in lung cancer patients from CT data. In addition to regularly contributing to open-source projects including modelhub.ai and PyRadiomics, he is also intrigued by data visualization, web development, and everything UI/UX. He has previously conducted research at Brigham and Women’s hospital, Wyss Institute for Biologically Inspired Engineering, as well as MIT Media Lab’s Mediated Matter group. As an architect and computational designer in a former life, he spent 4 years working in construction with Foster+Partners in Beijing and Playze in Shanghai. He is currently working on a PhD in machine learning and medical imaging at Maastricht University after having completed a Master of Design Studies in Technology at the Harvard Graduate School of Design, and a Bachelor of Architecture at the American University of Sharjah.

Bill Lotter is the CTO and a co-founder of DeepHeath, whose aim is to bring the best doctor in the world to every patient, through machine learning. Their first domain of focus is screening mammography, where they have developed award-winning machine learning models. Bill has a PhD in Biophysics and Computational Science from Harvard University. His thesis focused on biologically-inspired deep learning methods. Prior to Harvard, he worked as an algorithmic trader and also has served as a machine learning consultant for an NFL franchise.

Tal Schuster is a PhD student in Computer Science at MIT and a member of the Learning to Cure team. He is developing deep learning models that help reduce the radiologists’ workload by automating some of their routine tasks, improving risk assessment, detecting cancer earlier, and preventing overtreatment. Some of these methods are already successfully being used at Massachusetts General Hospital. Tal received his MSc in CS from Tel Aviv University and his BSc in Mathematics and CS from Ben-Gurion University. He had several key positions in the IDF, including leading a team of engineers working on automating procedures.

Liz Asai has served as CEO of 3Derm since 2013. 3Derm is a digital health company that has developed skin imaging systems paired with machine learning algorithms to triage dermatology concerns. Over the last few years, 3Derm has raised two rounds of funding, conducted three clinical trials, and obtained reimbursement for its dermatological services from several health plans. 3Derm now serves thousands of patients at health systems in the US. Liz holds a B.S. in Biomedical Engineering from Yale University and was featured in Forbes 30 Under 30.

Deep Learning Opportunities in Cancer Imaging

Radiographic medical images contain a vast wealth of information allowing for accurate non-invasive tumor characterization and ultimately improving cancer diagnosis and care. Recent advances in AI, deep learning in particular, promise to impact multiple facets of the radiology profession and support clinical decision making. The Computational Imaging and Bioinformatics Lab at Dana Farber Cancer Institute and Harvard Medical School is a data science lab focused on the development and application of AI methods on medical data.

In this talk, we will be presenting case studies investigating the clinical utility of deep learning in the detection, segmentation and outcome prediction of cancer tumors. We will also be presenting modelhub.ai, a repository of deep learning models crowdsourced through contributions by the scientific research community.

Ahmed Hosny is a machine learning research scientist focused on solving biomedical problems. Currently at Dana Farber Cancer institute, he trains and optimizes deep learning networks for the prognostication and treatment response prediction in lung cancer patients from CT data. In addition to regularly contributing to open-source projects including modelhub.ai and PyRadiomics, he is also intrigued by data visualization, web development, and everything UI/UX. He has previously conducted research at Brigham and Women’s hospital, Wyss Institute for Biologically Inspired Engineering, as well as MIT Media Lab’s Mediated Matter group. As an architect and computational designer in a former life, he spent 4 years working in construction with Foster+Partners in Beijing and Playze in Shanghai. He is currently working on a PhD in machine learning and medical imaging at Maastricht University after having completed a Master of Design Studies in Technology at the Harvard Graduate School of Design, and a Bachelor of Architecture at the American University of Sharjah.