Building a hospital referral expert system with a Prediction and OptimizationBased Decision Support System algorithm

Send This

Download

Comment

Favourite

more

Rating :

Rate It:

Embed :

Post a comment

Post Comment on Twitter Post Comment on SlideWorld

Comments:

Subscribe to follow-up comments

Notes

Hide Notes

Slide 1 :

Building a Hospital Referral Expert System with a Prediction and Optimization-Based Decision Support System Algorithm Chih-Lin Chi Health Informatics Program Joint work with Dr. Nick Street and Dr. Marcia Ward

Slide 2 :

Hospital Referral Tens of thousands Americans die each year and many more suffer from non-fatal injuries due to errors Health care quality varies across institutions Hospital-referral information are usually provided by physicians and advertisement It has been estimated that 2600 lives can be saved by hospital-referral each year

Slide 3 :

Quantifying Quality Outcome measure: measuring outcomes directly 30-day Risk Adjusted Death (Mortality) Rates Process measure: measuring the utilization of necessary interventions Aspirin and Beta Blocker at Arrival ACE Inhibitor or ARB for Left Ventricular Systolic Dysfunction Hospital characteristics: predictors of quality Discharge or surgery volume, bed size, teaching status, JCAHO accreditation, …etc

Slide 4 :

Problems behind These Approaches Certain hospitals may be over-utilized while others are under-utilized Evidence shows patients prefer local higher-risk hospitals over traveling to lower-risk ones May need to consider a patient’s risk and feeling Multiple factors: survival, distance to a hospital, patients’ factors, geographic area, other quality indicators (complication, LOS, etc) The trade-off question: should an 70-year-old AMI patient with diabetes choose a mid-size teaching hospital with JCAHO accreditation 20 miles away or a large and non-teaching hospital 30 miles away?

Slide 5 :

The individualized hospital referral Hospitals A, B, C, and D Individualized risk estimation for patient X X B 94% D 93% C 95% A 90%

Slide 6 :

The Hospital-Referral Decision Considering Several Factors A human expert’s hospital referral: online information rely on communication The ideal hospital recommendation: communication and optimization Customized hospital referral => expert systems

Slide 7 :

Motivation Individualized referral Individualized hospital-referral decision support The hospital that reduce death and suffering with reasonable traveling time The construction of expert systems Automating construction and maintenance of expert-systems using classifiers and optimization

Slide 8 :

Expert Systems A computer program that can make inferences and give conclusions using the knowledge of specialists. Clinical Decision Support Systems (CDSS): risk of a patient, pathology, the effect of a treatment plan, drug-interaction, insurance, etc. Query (symptoms, test results, etc.) Expert Systems (CDSS) Knowledge Source Inferences (e.g., treatment plan)

Slide 9 :

Current Expert Systems: Knowledge-Base Systems (KBS) Efforts of building KBS Knowledge source: literature review or experience of physicians Construction of the knowledge base: rules, graphs, trees, or networks Giving recommendations: inference engine, e.g., Bayesian Network The knowledge of KBS is transparent

Slide 10 :

Current Expert Systems: Case-Based Reasoning (CBR) Instead of human-derived knowledge, the knowledge of CBR comes from previously solved cases which are stored in case library A case comprises the problem the solution the outcome Giving recommendations: finding the case with the most similar problem to the query

Slide 11 :

Prediction and Optimization-Based Decision Support System (PODSS) Algorithm Knowledge source: generated from supervised machine learning methods, such as Support Vector Machines Inference generation: optimization methods

Slide 12 :

The Comparison of Expert Systems Both CBR and PODSS can avoid the bottleneck problem Knowledge source: human expert (KBS), data (CBR), and functions derived from data (PODSS) Clinical trial studies Literature reviews Knowledge base Coding rules Executive engine Case library Similarity Evaluation function Optimization KBS CBR PODSS Knowledge source Inference generation Bottleneck problem

Slide 13 :

The Hospital-Referral Expert System Query: a patient’s characteristics and the distance limit Objective function: highest survival and/or freedom-from-complication (FFC) rate Constraints: traveling distance limit Query PODSS The hospital with the highest survival rate The hospital with the highest survival rate and highest FFC rate Single-objective problem Multi-objective problem

Slide 14 :

Datasets The 2004 State Inpatient Dataset (SID) for Iowa from the Agency for Healthcare Research and Quality (AHRQ) Healthcare Cost and Utilization Project (HCUP) American Hospital Association (AHA) Acute myocardial infarction (AMI) with ICD-9-CM codes of 410.01 to 410.91 116 nonfederal acute-care hospitals in Iowa Descriptive variables and zip codes of both patients and hospitals Euclidean distance estimation based on two zip codes

Slide 15 :

Datasets Two types of desired outcomes: survival and freedom-from-complication (FFC) Datasets for each problem Single-objective optimization (survival) All AMI patients AMI patients without any surgery AMI patients with only coronary artery bypass graft (CABG) surgeries (ICD-9 36.10 to 36.19) Multi-objective optimization (survival + FFC) AMI patients with only CABG surgeries

Slide 16 :

Support Vector Machines (SVM)Classifiers Finding the decision function (e.g., d(x)=wx-b) that can classify cases with minimum error and maximum margin Kernel trick to facilitate using non-linear decision functions Using RBF kernel +++ +++ +++ +++ +++ +++ +++++++ + ++ +++ ++++ +++ ++++++ + ++ + ++++ ++ + ++++ ---- ----- --- ------ ------ --- ----- -- --- -- - -- ----- ---- -------- d1* d2

Slide 17 :

Maximizing the Desired Outcome Probability Maximize the desired-outcome probabilities, P(Y=1|X) or d(X) Y is binary outcome 1 (desired outcome, e.g., survival) -1 (undesired outcome, e.g., death) X are independent variables (X1 U X2j) X1: uncontrollable variables or constant: the query, e.g., a patient’s characteristics X2j: controllable variables or decision variables: actions that influence outcomes, e.g., hospitals Objective function: max d(X1 U X2j)

Slide 18 :

Maximizing the Desired Outcome Probability A query, X1, and three hospital-staying results: A*, A+, and A- P(Y=1|A*) > P(Y=1|A+) > P(Y=1|A-) Maximizing the confidence toward the desired outcome Goal: finding A* and the optimal action (X2*) decease survival

Slide 19 :

Problem Types and Optimizers Expert system problems: deciding X2j Construction problem: constructing a solution from scratch under certain constraints (e.g., drug type and the dosage) Selection problem: selecting the best one from possible solutions (e.g., hospital selection) 116 nonfederal acute-care hospitals Each hospital has a hospital descriptive vector, X2, {bed size, metropolitan status, discharge volume, etc}

Slide 20 :

The single-objective optimization The hospital with the highest survival probability Formulation Finding the hospital vector by maximize survival rate subjected to the traveling distance limit to a hospital Communication: The query: patient’s characteristics, X1, and distance limit, DL The expert system: return the optimal decision variables (the optimal hospital), X2*, and the optimum survival rate d*

Slide 21 :

The multi-objective optimization The hospital with the highest survival and FFC probabilities Formulation Maximizing survival (d1), maximizing FFC (d2) and minimizing distance (d3) The balance of weights of the three targets should be decided by users

Slide 22 :

Constraints and Formulation The location of decision functions may vary Cost-effectiveness: minimized cost with acceptable effectiveness (e.g., survival>=95%) Using ML to capture uncertain information as the decision function (e.g., effectiveness) Incorporating the knowledge of domain experts Traveling distance-limit (the risk of traveling) from query Drug-interaction in treatment recommendation Range of nutrition amount in life-style selection problem

Slide 23 :

Validation Map Calibrate decision scores into probabilities (Platt’s calibration) The validation problem Clinical trial Training parameters with independent calibration set Storing validation information (unbiased estimate of survival rate) Improvement of decision scores vs. probabilities

Slide 24 :

Validation Map Hospitals that can perform CABG surgery are usually large

Slide 25 :

Data Distribution Over-sampling of the minority class

Slide 26 :

Single-Objective Optimization

Slide 27 :

Comparison between True and Predicted Survival Probabilities

Slide 28 :

Multi-Objective Optimization: a Case Study The user decides the balance of weights and find the solution Efficient presentation of the solution space

Slide 29 :

Aggregated results

Slide 30 :

Other Results Three instead of one globally best hospitals Nonlinear effect Unique advantages of the three hospitals Can identify small but good hospitals in rural area Not always the larger the better

Slide 31 :

Conclusion Customized hospital referral decision Prediction- and optimization-based expert system Capture or update knowledge by classifiers and conclude an inference by optimization methods The recommendation is an action with the maximum confidence to the desired outcome May incorporate knowledge from domain experts

Slide 32 :

Future Work Multi-class problems Ranking problems Validation method Treatment recommendation Life style recommendation

SUPPORTING CAREER OP...

Strategic Management...

Vital Signs A Practi...

Serial Murder in Hos...

Caseview building th...

Prediction of pulmon...

Free Powerpoint Templates

	chih-lin-chi@uiowa.edu 5 Years ago.



2343 Views, 0 favourite
PowerPoint Slide Presentation on
More By User Flag as inappropriate Select your reason for flagging this presentation as inappropriate. If needed, use the feedback form to let us know more details. Cancel