Randall Davis

We want to demonstrate that multimodal analysis of simultaneous hand and eye movements can both measure a subject’s overt behavior and provide new insights into how they are thinking i.e., their methods and strategies. Insights into strategies in turn provide opportunity for early diagnosis and early intervention for both neurodegenerative disorders in adults and neurodevelopmental disorders in children. The proposed work is the next step an ongoing project exploring grapho-motor/cognitive interactions that has been active for more than a decade. The new work adds additional modalities that provide exceptional opportunities for multimodal interpretation. We have been working on digital cognitive biomarkers for more than a decade, developing novel tests that measure cognition quickly, inexpensively, and with the objectivity and repeatability that comes from high resolution measurements and software-defined analytics. Within the larger scope of this ongoing work, the primary focus of the proposed work is the collection, cleaning, ML analysis, and interpretation of multimodal (drawing motion and eye motion) subject data.

Rosalind W. Picard

The disease of Major Depressive Disorder (MDD) is one of the worldwide leading causes of disability. At MIT, approximately 22% of undergraduates suffer from severe or moderately severe depression. Depression is the main contributor to suicide, and the US has seen suicide rates increase in nearly every state over the period from 1999- 2016, with half of the states having rate increases of more than 30%. MDD is associated with behavior, mood, sleep and physiological alterations, which can be measured with multi-modal sensors from wearables and smartphones. Studies have shown that multimodal and physiological data from mobile and wearable devices can be used to reliably estimate depression severity assessed by psychiatrists, detecting depressed and euthymic (remission) phases. Today’s methods have significant limitations that we will address in this proposal by developing novel personalized machine learning techniques for monitoring disease progression and enabling personalized medicine. Our project goals are to develop new personalized Machine Learning (ML) approaches to accurately monitor changes in MDD level and to forecast illness trajectory, identifying the most influential variables and enabling personalized time-adaptive treatments. We have access to clinical data and are working closely with the #1-rated-psychiatry in the USA, MGH, to shape its clinical impact. Our research will facilitate the implementation of personalized digital medicine by allowing prescription of specific therapy on the basis of the forecasted disease profile identified from passive data, and ultimately increase the effectiveness of selective prevention strategies. This will enable delivery of optimal MDD prevention or treatment to those who currently cannot access it due to time constraints, lack of available services, trust, or cost. This work also will provide new advances in AI/machine learning, which we will disseminate through publications, media, and presentations. Our default plan is also to open source our personalized machine learning software, to expedite advances across both AI and personalized medicine.

Muriel Medard

Machine learning (ML) algorithms have attracted significant attention for classification tasks for many sensitive medical purposes, e.g., diagnosis, predictions, etc. Training a ML classification model with a high precision for medical purposes requires advanced computing resources and sufficient training data collected from diverse groups of patients, e.g., age-wise, ethnicity-wise, etc. In practice, sufficient and diverse training samples do not exist at the same time and location. Thus, many medical centers would like to collaboratively train ML models on their combined datasets for a common benefit. Moreover, advanced computing resources and software (e.g., GPUs, TPUs, CuDDN) are costly and do not usually exist at every medical center. An attractive solution is to outsource the training data from several centers to a cloud server where the ML model is trained. Cloud providers begin to offer Machine Learning as a Service (MLaaS), a new service paradigm that uses cloud infrastructures to train models and offer online prediction services to clients, e.g., Microsoft Azure, Google Cloud AI. However, sending plaintext sensitive data to third parties for processing is highly undesired due to privacy concerns and/or business competition, e.g., see General Data Protection Regulation (GDPR) and Health Insurance Portability and Accountability Act of 1996 (HIPAA). The proposed coding scheme allows multiple parties to train a classification ML without revealing their data. The intuition behind the proposed method is encrypting training data at each medical center via a low-complexity coding scheme, and transferring the encrypted data to the data center and training the ML model with encrypted data rather than original raw data. Our goal is to present a solution that offers encryption of the sensitive datasets such that an information-theoretic privacy bound can be guaranteed, and thus no attacker can possibly infer information about the original data from the encrypted data or from the trained model beyond a certain limit. The model that is trained with encrypted data then can be used for inference at the data center or be transferred back to the local offices. In the proposed scheme, each medical center has its own encryption scheme, aka security key, that is computationally very difficult to be guessed and does not need to be shared with other centers, organizations, cloud, etc. Moreover, since the ML model is trained with the encrypted data, only authorized users who have access to a valid security key (the local offices that provided training data) can use the devised classifier. We propose to employ Random MultiLayer Perceptron (RMLP) networks to convert original sensitive samples into a new representation that provides privacy.

Vivek Farias

Recent years have seen an explosion in the availability of large-scale observational datasets where the questions at hand may be viewed as ones of causal inference. As a concrete example, consider the Connectivity Map (CMap) datasets. These datasets describe the cellular impact – expressed in terms of genomic (or proteomic) expression – of systematic perturbations by a large set of genetic and pharmacologic perturbagens. For instance, CMap contains approximately 1.5M gene expression profiles across approximately 100 cell lines perturbed by approximately 5,000 small molecule compounds. This data was made possible by a relatively inexpensive, high throughput gene expression profiling technology dubbed L1000. In a similar vain, the PI has recently helped develop a high throughput platform for unbiased proteomic sampling which has the potential to provide similar data for unbiased proteomic profiling. This platform has already shown promise as a tool to generate proteomic biomarkers at scale for both non-small-cell lung cancer (NSCLC) and Alzheimers disease.

Sixian You

Optimizing the imaging design strategy within the constraint of low-resource clinical environments is crucial for achieving accurate cancer detection despite the severely constrained environment. This project aims to develop an AI-based framework that learns an illumination/detection strategy for optimized optical imaging system design within the constraint of point-of-procedure breast cancer imaging1. The optical imaging system (illumination and detection scheme) will be jointly optimized with a neural- network-based tumor detection method in an end-to-end (image generation to image interpretation) fashion for efficient co-design.

Kalyan Veeramachaneni

Our mission is to create a digitized early detection cognitive assessment tool that will provide accessible, noninvasive, scalable assessments that are quick and comprehensive. Currently, a combination of multiple diagnostic tools is required to pinpoint cases of cognitive impairment. These tests have reasonable specificity and sensitivity, but require a significant time to complete and cannot be used as routine screening exams. 1 The current research standard, the amyloid scan, provides a presymptomatic diagnosis, but it only has a sensitivity of 90% and specificity of 85%, 2 costs over $3000, and is not commercially available. Generally, physicians must rely on patient history, family input, physical exams, and cognitive tests to make a clinical diagnosis of this condition. While there is no known cure for dementia, treatments can slow its progression. We developed iBOCA, an accessible, culturally- competent app, that allows users to track their disease progression and clinicians to monitor their patients’ cognitive health. By providing a single, comprehensive tool for early detection, treatments can delay cognitive decline and improve overall quality of life. iBOCA will combat the public health disparity caused by expensive diagnostic procedures and reveal new findings about cognitive impairment through fine granular data collection.

Polina Golland

In this project we will develop novel machine learning algorithms for accurate non-invasive assessment of pulmonary edema in patients with congestive heart failure from chest x-ray images. The proposed methods will jointly model images with other available data such as radiology reports and clinical indicators during learning to improve the image-based assessment at inference time. The resulting approach to image-based edema assessment will be applicable broadly beyond the applications to congestive heart failure that motivate our research. Our deployment plan includes integration into the hospital information systems at our clinical collaborator site and other institutions.

Peter Szolovits

Missing and conflicting data constitute an increasing risk to patient safety and adequate care. Alerts generated in the telehealth environment focus on surveillance of vital sign data and signals telehealth providers about parameters that exceed configurable thresholds or trends in physiological variables and specific multi-parameter combinations. For example, two consecutive average heart rate values above 150 beats per minute (bpm) will trigger a default alert as will a 10-minute median value above 130 bpm. When there are missing data, the alert may be delayed until enough complete data are available or future alerts may not occur at all. Further, the same patient health parameter may be measured by an electronic mechanism and by a human being. The latter is usually applied and the former is ignored. At one medical center, the difference has been found to be the largest in the cases of patients who soon die. We propose to leverage the concept of Bayesian networks to perform the task of data imputation using test data from MIMIC and other sources in the public domain.

David Sontag

This proposal seeks to develop machine learning methods to automatically retrieve contextually-relevant clinical data (e.g., notes, labs, imaging) from the patient’s medical history to improve clinical decision making at the point of care. This project is a continuation of a collaboration between MIT and Beth Israel Deaconess Medical Center, previously funded by Jameel Clinic, that has already resulted in the development of a new AI-driven user interface for clinical documentation that has been deployed in BIDMC Emergency Department.