Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Artificial Intelligence & Clinical Decision Support. Including fuzzy logic, neural nets, and genetic algorithms CSE 5810 Kevin Lopez Computer Science & Engineering Department The University of Connecticut 371 Fairfield Road, Storrs, CT 06269-2155 [email protected] Lopez-1 What is Clinical Decision Support? CSE 5810 Clinical Decision Support is: Knowledge provided to clinicians From Multiple Sources/Contexts, processed and returned in a form that will assist a care giver. Involves processing via various artificial intelligence and machine learning technologies CDS is Multi-Disciplinary Computing (Information Processing, Data Analysis) Social Science (User Interactions) Clinical Decision is applicable to many domains: Can be used in any type of medicine, including domains with weak domain theory. Its underlying systems (AI) is used for any field Lopez-2 What is Clinical Decision Support? CSE 5810 Key people(s) affected: Patients Physicians, clinicians, care givers Hospitals/medical centers Standards: Arden Syntax (Syntax) GELLO (Common Expression Language) Infobutton (Context-aware Knowledge Retrieval) Techniques: These are still being worked on and researched. No set technique Lopez-3 What is Clinical Decision Support? CSE 5810 A combination of different knowledge's. Knowledgebase (Textbook, etc) Clinicians Knowledge/experien ce Gained Experience from learning, and individual patients Clinical Decision Support System Knowledgebase Clinician Gained Experience Lopez-4 Why use a CDSS? CSE 5810 We use a CDSS because of: It provides better quality of care Can provide the clinician with a second opinion Can guide a novice clinician to a solution, diagnosis, or treatment. Can help reduce the number of errors It can help with the speed and quality of diagnosis It improves customer/patient satisfaction Can be interactive (with the clinician) to get the best results. Can be nearly autonomous, some systems are personal and can give a diagnosis. Lopez-5 Functions of a CDSS CSE 5810 A CDSS generally works by: Taking in some data, normally it is some patient data This data can be measurements, clinician data, or knowledgebase data. The data then must be extrapolated and the most relevant parts used for processing. The data is then processed with the method of choice (ANN, CBR, Fuzzy etc.) and may require clinician input as well. The data is then post processed and outputted in a variety of fashions (can be numerical, binary, or even text). Lopez-6 Designing a CDSS CSE 5810 Main problems these systems must solve Structured These problems are routine and repetitive Solutions exist, and are standard and predefined Unstructured Complex and fuzzy Lack Clear and straightforward solutions Semi-structured This is a combination of the two previous catagories. Lopez-7 Artificial intelligence's role in clinical decision support CSE 5810 Two types of CDSS Work with Knowledgebase Work with Non-Knowledgebase Knowledge based CDSS: Use knowledge from sources such as textbooks, and other resources. They have rules similar to if-then statements. Components of a knowledge based CDSS: Knowledgebase: Some source where they get their knowledge Inference engine: takes data and applies the rules from the knowledgebase Communication: Allows system to communicate with user and user input. Lopez-14 Hybrid Systems CSE 5810 Hybrid systems Knowledge and Non-Knowledge based system These systems produce high quality results from the merge of the two different systems. They have an already established knowledge base but they also must learn from past experiences or from test results. These systems often Produce results that are better than these systems individually. These systems can be a combination of many of the different technologies that each system has. Lopez-16 Artificial Neural Networks CSE 5810 Similar to real neural networks Take in data and pass them through the network to the other neurons to get an output. Many times used for pattern recognition Several different algorithms can be used for threshold Lopez-18 Case-Based Reasoning CSE 5810 Case-based reasoning is: A process of solving new problems based off of old problems. Similar to how humans think and solve problems. Can take new solutions that have been solved and add them to the database of solutions for future reference. There are Four Steps (R’s) to case based reasoning: Retrieve: where the system retrieves the knowledge Reuse: takes old experience and maps it to new problem Revise: revise the solution Retain: put new solution into the system database Lopez-22 Case-Based Reasoning The four R’s for Case based reasoning CSE 5810 Lopez-27 Fuzzy Techniques CSE 5810 Fuzzy Logic is: Degrees of truth, 0 and 1 are extremes. Some types of data do not have what we consider a full truth or false. An example of Fuzzy Logic An example of this is natural language processing. This is where truths are aggregated from partial truths. This is to derive meaning from humans such as notes a doctor put in or some other source of natural language. Lopez-28 Genetic Algorithms CSE 5810 Based off of a simplified evolutionary process used to arrive at an optimal solution. It works in the following way: Children are made and try to solve the problem The top few children then are used to generate new children This process continues until an optimal (or very close to optimal) solution is found. In CDSS: The selected algorithms evaluate the solution Of these solutions the best are chosen and they try to evaluate the problem again until the solution is found. Lopez-30 Feature Selection CSE 5810 Feature Selection is: Selecting features or attributes from a set of data Useful for taking out certain data that is not needed during processing Similar to how we process data, we do not need to know all of the data but we extract key items from the data. Data may have redundant features that provide no more information as the features previously selected. Feature Selection is used in getting the data that is required. Allows for less and unnecessary processing. Lopez-31 Personal Medicine CSE 5810 There are several apps that claim to assist with diagnosis. In particular several skin cancer apps have surfaced. None of which are free Some of which incorporate sending the images to a clinician for further diagnosis. Some of the apps have the ability to use the camera to view the skin and take a picture With this picture the program checks for symptoms, or “ugly duckling moles” Apps are still improving to give more quality care Lopez-35 Personal Medicine CSE 5810 Lopez-36 Effectiveness of CDSS CSE 5810 How effective are these systems CDSS’s are becoming more and more effective and accurate at diagnosing diseases. Many times these systems improve the outcome of both treatments and diagnosis of patients Many times these systems are integrated into the clinicians workflow to provide superior satisfaction to both the patient and the clinician. These systems give the clinician a recommendation not just an assessment, so that the clinician can actually follow through. These systems many times outperform their clinician counterparts in diagnosing a patient. Lopez-39 Key Technical Problems CSE 5810 Some of the problems that are seen with CDSS Many different types of artificial intelligence that serve many different purposes No one generic algorithm that can handle all of the data Natural language can be very difficult to extract data from Some domains have weak domain theory Many of the systems need time to train and much of the training is computationally expensive Data preferred to be shortened (feature selection) in order to take less time processing. Lopez-42 Key People Problems CSE 5810 There are problems that exist where the user may experience either due to lack of experience or familiarity. Ease of use: The system must be easy to use, and work right out of the box. There should be minimal configuration if any done by the clinician. The interface has to be user friendly. Many times users of these systems have very little computer knowledge. The user should not have to be trained on this system. Data input: the data must be entered correctly (ie. switching systolic and diastolic). Lopez-43 Conclusion CSE 5810 These Systems Show: Improvement in patient outcome Higher Patient satisfaction Guidance for inexperienced practitioners Guidance for individuals These systems cannot: Replace a doctor/care giver Are limited in how many different diseases each one can do Be 100% accurate/fool proof Lopez-44 References CSE 5810 Application of Artificial Intelligence for Clinical Decision Making and Reasoning (Abdalla S.A.Mohamed) Efficient Clinical Decision Making by Learning from Missing Clinical Data (Farooq, Yang, Hussain, Huang, MacRae, Eckl, Slack) Developing Decision Support for Dialysis Treatment of Chronic Kidney Failure the researchers explore and describe what goes into developing a CDS system for dialysis treatment. Hybrid Case-Based System in Clinical Diagnosis and Treatment. A Model to Predict Limb Salvage in Severe Combat-related Open Calcaneus Fractures Clinical Decision support system for fetal Delivery using Artificial Neural Networks the team are using ANN’s to assist doctors with decisions at critical times of fetal deliveries. Implementing Decision Tree Fuzzy Rules in Clinical Decision Support System after Comparing with Fuzzy based and Neural Network based systems Case Studies on the Clinical Applications using Case-Based Reasoning Improving clinical practice using clinical decision support systems: a systematic review of trials to identify features critical to success (Kensaku Kawamoto, Caitlin A Houlihan, E Andrew Balas, David F Lobach) Effects of Computerized Clinical Decision Support Systems on Practitioner Performance and Patient Outcomes E-Health towards Ecumenical Framework for Personalized Medicine via Decision Support System Standards in Clinical Decision Support: Activities in Health Level Seven And Beyond (https://www.dchi.duke.edu/conferences/posters-presentations/amia/2011-amia/KawamotoStandardsInClinicalDecisionSupport_slides.pdf) Kai Goebel from Rensselaer Polytechnic Institute (http://www.cs.rpi.edu/courses/fall01/softcomputing/pdf/cbr1to3.pdf) HealthIT (http://www.healthit.gov/policy-researchers-implementers/clinical-decision-support-cds) Lopez-45