Comparative Effectiveness Research (CER) hypothesis prediction in - - PowerPoint PPT Presentation

Comparative Effectiveness Research (CER) hypothesis prediction in personal health messages

Yunliang Jiang (Computer Science, UIUC) Advisor: Bruce Schatz (Medical Information Science, UIUC) With: Vera Qingzi Liao (UIUC) and Qiaozhu Mei (Umich) Sponsor: USDA-NRI, UIUC-IGB, State Farm Scholarship

Approach

• Unstructured Message filtering, extraction
• Professional labeling by medical students
• Sentiment Analysis (attitude: pos/neg/neu)
• Direct comparison by same patient (e.g. I

prefer T1 to T2)

• Indirect comparison by same patient (e.g.

positive to T1, negative to T2)

• Indirect comparison in overall case
• Demographic Analysis (sex, age, region)
• Statistical test

Problem

• How to compare the effectiveness of different

drugs/treatments by personal health messages?

• Patient’s opinion on treatments (attitude, preference)

can interpret the effectiveness

• Data source: MedHelp (100+ forums, 1M+ messages)
• Representative Dataset:
• Breast Cancer (70K messages, 18K patients,

Chemo v.s. Radiation v.s. Hormonal)

• Depression (186K messages, 38K patients,

Meditation v.s. Drug treatment)

Conclusion and Plans

• Our approach can predict patients’

preference consistently (e.g. people prefer meditation to SSRI)

• Different groups may have specific

preferences (e.g. Age 65+ people prefer radiation than chemo; South people prefer Hormonal therapy more than Northeast people.)

• Predicted hypothesis => need clinic proof

Text Mining for Health Informatics

Parikshit Sondhi (UIUC)

Advisor: ChengXiang Zhai (UIUC) with Jimeng Sun (IBM), Hanghang Tong (IBM)

Problem

• Mining symptom relationships from

unstructured patient records

• Dataset:
• CHF Cases: 500K (4.6K Patients)
• Controls: 500K (8.4K Patients)
• Find symptom mentions related to

known CHF symptoms Results and Conclusion Approach

• Construct symptom relationship graphs
• Similarity via Random Walk Restart

Heart Attack Rales Cough Wheezing Chest Pain Fever Ankle Edema w1 w2 w3 w4 w5 Heart Attack Rales Cough Wheezing Chest Pain Fever Ankle Edema

Successfully identified related diseases and symptoms, confirmed by clinical experts

Siddharth Gupta (UIUC) Advisor: Carl A. Gunter (UIUC)

with Mario Frank (UC Berkeley) and SHARPS

Approach Problem Progress and Plans

• Insider threats
• curiosity and personal information
• financial fraud
• medical identity theft.
• Complex workflows and difficult to handle

access in emergency situations.

• Big Data with limited semantic knowledge

about user-patient interactions.

• Experience Based Access Management (EBAM)
• Cluster the users and patients in a high

dimensional space based on features such as (diagnosis, procedures, medications)

• Type all users based on interaction with the

patient clusters.

• Completed modeling (LDA) to form clusters of

users and patients.

• Selected 15-25 Topics for each feature, using

perplexity measure.

• Analyzed probability of a user belonging to a

patient topic and typing them.

• Plan to evaluate the model using “random

user model”.

Detecting Anomalous Accesses in EHR Audit Logs

Dataset: 7 million access logs, 8000 Users and 25000 Patients

Processing of raw data generated by phone sensors to extract walking motion and build gait detection models

Qian Cheng (Computer Science, UIUC) Advisor: Bruce Schatz With: Joshua Juen (Electrical and Computer Engineering, UIUC ), Yanen Li (Computer Science, UIUC), Yunliang Jiang ((Computer Science, UIUC)

Sponsor: USDA-NRI, UIUC-IGB

Approach Feature extraction

• Sampling frequency: 60Hz
• time domain features: mean, standard

deviation and mean crossing rate

• frequency domain: peak location, entropy,

sub-band energy and sub-band energy ratio Training data

• 5min walk for each type of motions; carry

phone in different place; user feedback to label each set Server operation

• Receive and storage all useful data
• Run models and return detecting decisions

Problem

• Raw data to gait features
• Personalized model for detecting abnormal

health

• Integrated system with phone clients and

servers Progress and Plans Time-series data process

• Combine the acceleration and orientation
• information. Interpret acceleration in body

based vertical-horizontal coordinates

• Iteratively split data to same pieces(2000pts)

Frequency domain data process

• Use FFT to interpret time-series data to

frequency domain.

• Filters to remove noise generated by phone

Model building

• Decision tree and SVM for classification
• Semi-supervised strategy to improve models

Title: Selective Consumption of Online Health Information

• Q. Vera Liao (HCI Group, CS Dept.)

Advisor: Prof. Wai-Tat Fu Approach Problem:

• Internet provides overwhelmingly large amount,

diverse, and sometimes contradicting information. to health information seekers

• Varying quality: malicious source, unreliable

information

• Varying opinions: Information bias
• I am interested in understanding and designing

system/interface in support of health information seekers’

• Selection of high quality information
• Deliberation on diverse opinions

Progress and Plans

• Experiment 1: credibility judgment
• Experiment 2: selective exposure bias
• In progress: analyze data on Medhelp.com to see

how users interact with different opinions.

• Plan to conduct more user studies to understand

how to nudge users to make “better” selection

• Plan to develop system or interface that supports

users to navigate through the diverse opinion space

• f online health information when facing complex

health related decisions

Empirical studies to analyze user behavior patterns Case studies on real health info websites Design “informed” system and interface

Building 3D Tele-immersive (3DTI) Applications for Remote Physical Therapy Approach Problem Progress and Plans

• - Remote physical therapy requirements
• - Definition of 3DTI system
• 1. High Bandwidth usage
• 2. Low delay for interactivity
• - Deficiency of current technology
• 1. 3D video acquisition, compression, transmission
• 2. Multi-stream synchronization
• 3. Intuitive user-interaction (UI) solution
• - Current:
• 1. A working 3DTI testbed
• 2. 3D semantic-based middleware implementation
• 3. TEEVE-Remote UI
• - Future:
• 1. Explore the multi-modality from the perspective
• f synchronization
• - Real-time 3D data acquisition and model

reconstruction using Kinect camera

• - 3D semantic-based compression and

transmission framework

• - Synchronization framework for multi-modal

3DTI streams

• - Intuitive UI solution using mobile phones and

Kinect camera Henry Pengye Xia Advisor: Klara Nahrstedt TEEVE@UIUC

Approach

Problem

Progress and Plans

Safe Reconfiguration of Acute Care Monitoring Systems

Maryam Rahmaniheris (UIUC)

Advisor: Lui Sha(UIUC) Collaborator: Richard Berlin (UIUC)

• Medical Device ”Plug-and-Play” (MDPnP) can

mitigate the preventable medical errors

• Dynamic clinical environment poses new design

challenges

• The following questions must be answered
• Is the current configuration safe and consistent?
• Did the system reconfiguration cause changes in the

measurements interpretation?

• An abstract model of dynamic clinical environment
• A highly modular system architecture
• Encapsulation of MDPnP complexity
• Consistent propagation of necessary updates

throughout the monitoring system

• An abstract model of monitoring system is proposed
• An MDPnP architecture with proven safety and

consistency is developed

• Performing more thorough analysis of the architecture
• Collecting more use cases to evaluate the efficiency of
• ur proposed model and architecture to enforce

safety and consistency in different acute care scenarios

Biosensor for Point of Care HIV Diagnostics for Global Health

Approach Progress & Results Device & Future Directions

Umer Hassan (UIUC), Advisor: Rashid Bashir (UIUC)

The biggest challenge the country faces today is diagnosing all of its HIV-infected people and helping them take full advantage of the existing treatments (Science, p.167, Vol. 337, 13 July 2012).

Problem

y = 0.9814x R² = 0.8772 400 500 600 700 800 900 1000 1100 1200 400 500 600 700 800 900 1000 1100 1200

Chip CD4 T Counts /uL Carle Control CD4 T Counts/uL

The CD4 Counts from our device shows a good correlation with the control counts

• btained from the Carle Hospital

The Differential counting of CD4 T lymphocytes using the whole blood

• Start testing the current device on the patient samples
• CD4/CD8 Counting on the same device
• Integrating flow metering in the device

Pillars in Capture Chamber CD4 Counter Device with Flat Chambers

Proposed Plans

Body Area Network Security: Robust Key Establishment Using Human Body Channel

Sang-Yoon Chang (UIUC ECE), Advisor: Yih-Chun Hu (UIUC ECE), with Hans Anderson (UIUC ECE), Ting Fu (UIUC MCB), Evelyn Y. L. Huang (UIC Medicine) (Based on work in HealthSec 2012)

Approach

∎ Establish secure channel using body-coupled communication using artificial signal ∎ Operate below action potential ∎ Built our scheme and demonstrate successful signal transfer on dead mouse ∎ To simulate living human body medium:

• Study signal attenuation with

homogeneous meat and dead mouse

• Noise measurement on

human subject

Motivation

∎ Communication within a body area sensor network, e.g., implantable/wearable sensors ∎ Security against outsiders ∎ Current key sharing schemes using natural biometrics (ECG) are not robust to sensor deployment location and vulnerable against outsider eavesdropper (side channel)

Progress and Plans

∎ Conservative and location-robust lower bound on body channel capacity of 11 bits per day (tolerable performance since secret updates only necessary when change in network infrastructure) ∎ More precise performance analysis for future work ∎ Also interested in other robustness and security issues with body sensors

Genomics and Privacy

Muhammad Naveed (UIUC)

Advisor: Carl Gunter (UIUC) with Xiaofeng Wang (IU), Haixu Tang (IU) Human Genome -> 3 billion base pairs

Approach

• Cryptography + Genomic Information
• State of the art cryptography does not

scale to genomic data

• Dramatic performance gains possible by

using domain knowledge about genomics

• Aim for integrated view with phenotype

data Problem

• Sequencing cost is decreasing
• Ever increasing amounts of intrinsically

identifying genomic data are available

• Sharing genomic data is useful
• How can we share genomic data in

clinical practice and research with privacy and security guarantees?

• Report to the president urges attention to

specified requirements Progress and Plans

• Mayo IM Conference: is genomic data

exceptional?

• Working with HPCBio Lab to get first

hand experience with genomic data

• Working on privacy preserving clinical

architecture for “genetically and symptomatically similar patient queries”

• Preliminary Results: 0.5 sec/comparison, for

20k SNPs disease marker

Crowdsourced Water Quality Monitoring Tristan Wietsma, INFO

• Dr. Logan Liu, ECE

Manas Gartia, ECE Approach Inexpensive, nanostructured, electro-chemical sensor for mobile broadband devices. Cloud-based infrastructure for data storage, sharing, and analytics. Problem It is difficult to measure water quality & safety. Even when accurate data is available, understanding the spatial characteristics is challenging. Progress and Plans Electrode developed for anions; calibrated for nitrate. In R&D for heavy metals, organics, and bacteria. Exploring cyberinfrastructure

• ptions.