Kansas State UniversityComputing and Information Sciences
IEEE CBMS ConferenceAugust 3-4, 2009
Speech Assisted Radiology System for Retrieval, Reporting and Annotaiton
Tim Weninger, Daniel Greene, Jack Hart, William H. Hsu and Surya Ramachandran*
Department of Computing and Information SciencesKansas State University, Manhattan KS
*AIdentity Matrix Inc, Elmhurst, IL
2009 IEEE International Symposium on Computer-Based Medical SystemsAlbuquerque, NM, USA
Kansas State UniversityComputing and Information Sciences
IEEE CBMS ConferenceAugust 3-4, 2009
Outline
• Introduction› Motivation› Example
• Voice Directed Search› Prerequisites› Parsing Spoken Text› Search› Findings and Impressions
• Merit Case Client› Experiments
› Metrics› Results
• Conclusions and Future Work• Demo
Kansas State UniversityComputing and Information Sciences
IEEE CBMS ConferenceAugust 3-4, 2009
Introduction
• Motivation› Paradigm: Radiology› Healthcare is expensive
› Why?
• Errors› 2004-2006 Medicare study
› Errors cost US$8.8 billon› University of Baylor study:
› Out of 113 errors studied› Transcription was the base-cause for 46%
› (Seely et al. 2004)
• Inefficiencies› Medical Transcription
› Adds cost› Adds complexity
Kansas State UniversityComputing and Information Sciences
IEEE CBMS ConferenceAugust 3-4, 2009
Introduction
• Status quo (simplified)
Kansas State UniversityComputing and Information Sciences
IEEE CBMS ConferenceAugust 3-4, 2009
Introduction
• Status quo (simplified)
Kansas State UniversityComputing and Information Sciences
IEEE CBMS ConferenceAugust 3-4, 2009
Definitions
• Define our Terms:› Paradigm: Radiology
› MRI› Finding/Impression
› Medical diagnostic interpretation of particular abnormalities as seen by the radiologist
› Annotation› The expression of a medical opinion related to a specific image.
› Drawn Arrow› Circle› Etc
• Merit Case Client:› Speech directed PACS system
Kansas State UniversityComputing and Information Sciences
IEEE CBMS ConferenceAugust 3-4, 2009
Voice Directed Search
• Current PACS systems› Example: find men with slipped discs
› “Search. sex equals male. diagnosis equals herniated disc.”› [Search] is a command› [male] is a menu option in list [sex]› [Herniated disc] is option in list [diagnosis]
• Disadvantages:› Narrow speech scope› Voice recognition systems are not foolproof
› Example: Homonyms› “Search. Sex equals mail. Diagnosis equals herniated disk.”
› Does not compute!
• Main advantage:› Capable of standardizing naturally spoken medical terminologies with
significant degrees of variance.
Kansas State UniversityComputing and Information Sciences
IEEE CBMS ConferenceAugust 3-4, 2009
Voice Directed Search› Example:
› Find all male patients between the ages of 55 and 60 with a slipped disc in the L4/L5 region with no previous history of disc injury.
› “Find men with a slipped disc in the L4/L5 region”› [Find] is a command along with others› [male] is a interpreted to be [male] within [sex]› [between 55 and 60] is [55-60] within [age]› [slipped disc] is interpreted to be [herniated disc] within [disease]› [disc injury] looks for any [disc] within [disease]
• Moreover:› This widens the search scope› Voice recognition systems are not foolproof
› Example: Homonyms and formatting› “Find men with a slipped disk in the El four slash El five region.”
› This works as well.
How?
Kansas State UniversityComputing and Information Sciences
IEEE CBMS ConferenceAugust 3-4, 2009
Parsing Spoken Text
• Operating Assumptions:› The system maintains a complete list of all ages, sexes, diseases, etc.
i.e. type enumeration› Valid responses are available in lists
› Homonyms do not coexist in a list› If so, then it’s hard to make a decision
• Goal› Map what is dictated to the appropriate descriptor
• Sliding window approach:There is moderate disc bulging at L5/S1
SizeSmallSmall to ModerateModerateModerate to LargeLarge
There is moderate disc bulging at L5/S1
DiagnosisDisc bulgingHerniated discDegenerative disc disease
Kansas State UniversityComputing and Information Sciences
IEEE CBMS ConferenceAugust 3-4, 2009
Voice Directed Search
• Synonym Learning› How does the system know:
› “Slipped Disc” = “Herniated Disc” = “Disc Herniation” › The system will make an initial guess.
› System will not initially recognize “Slipped Disc”
• System remembers corrections› Correction process is easy› Learns speakers word choice preference
Kansas State UniversityComputing and Information Sciences
IEEE CBMS ConferenceAugust 3-4, 2009
Structured Reporting
• Image Embedding› Findings› Impressions› Annotations
• Text, descriptors, drawings› Become linked with the image(s)
Kansas State UniversityComputing and Information Sciences
IEEE CBMS ConferenceAugust 3-4, 2009
Experiments
• Data points› (1) Text read by the radiologist› (2) Text output by speech recognition engine› (3) Descriptors filled in by Merit Case Client› (4) Correct state of the descriptor (ground truth)
• Metrics› Speech Recognition Metric (SRM)
› Word-Edit distance between original text (1) and output by the speech recognition system (2).
› Parsing Engine Metric (PEM)› Word-Edit distance between menus filled in by Merit Case Client (3) and
the correct answer (4)
Kansas State UniversityComputing and Information Sciences
IEEE CBMS ConferenceAugust 3-4, 2009
Experiments
• Reporting and Analysis› Some errors are more costly than others› 3 reporting methods:
› Word distance› Weighted errors
› Disease descriptor= 60%› Location descriptor = 20%› All others descriptors = 20%
› All or not› Was it completely correct or not?
• Experiment› Radiologist (Dr. Schekall, MD) made 100 dictations based on real-
world cases› 25 search queries› 75 findings and impressions dictations
› No re-dos allowed› Speech recognition system was NOT pre-trained
Kansas State UniversityComputing and Information Sciences
IEEE CBMS ConferenceAugust 3-4, 2009
Results
Data points and their linear regression lines
Kansas State UniversityComputing and Information Sciences
IEEE CBMS ConferenceAugust 3-4, 2009
Results
Change in accuracy for each paradigmMethod: (SRM-PEM)/SRM
Paradigm Test Accuracy (area) Change in Accuracy
Distance Speech (SRM) 87.65 +10.0285%
Parsing (PRM) 96.44
Weighted Speech (SRM) 66.93 +41.8945%
Parsing (PRM) 94.97
All-or-Not Speech (SRM) 35.81 +96.7607%
Parsing (PRM) 70.46
Kansas State UniversityComputing and Information Sciences
IEEE CBMS ConferenceAugust 3-4, 2009
Current domains of implementation (ongoing)
Branded under - Virtual Integrity in Medicine TM (VIM)
• Electronic Medical Records› VIM Radiology
› PET, CT, MRI, Nuclear, X-Ray, Ultrasound, etc› VIM Cardiology
› ECG, Ultrasound, CT, Nuclear, Cath lab, Vitals, Resting, Exercise, Stress, Ambulatory BP and Spirometry
› VIM Neurology› From out-patient clinical through surgery
• Front & Back Office› Scheduling, Patient profile, Insurance, Rule-outs ICD9/10, Referring Physician,
Reporting, Billing & Accounts bridge, Clinical messaging, etc.
Kansas State UniversityComputing and Information Sciences
IEEE CBMS ConferenceAugust 3-4, 2009
Demo
Kansas State UniversityComputing and Information Sciences
IEEE CBMS ConferenceAugust 3-4, 2009
Questions?
Special thanks -
Dr. Michael Schekall, MDDeborah Templeton, BS, CNMT, RT(R), LRT
Hutchinson Clinic PA, Hutchinson, KS
Jeff Barber, Andrew WaltersKansas State University
Industry Contact for more information –
Surya RamachandranAIdentity Matrix Medical [email protected]