Transcript PowerPoint Presentation - Computer Science & Engineering

A Causal Probabilitic Network for
Optimal Treatment of Bachterial
Infections
Alicia Ruvinsky
Scott Langevin
UNIVERSITY OF SOUTH CAROLINA
Department of Computer Science and Engineering
Problem: Bacterial Infections
• 30 percent mortality rate from severe bacterial
infection
• 1/3 given inappropriate treatment
• 20% prescribed superfluous drugs
• Anti-biotic drugs account for 20-50 percent of
hospitals drug expenses
• Bacterial resistance to anti-biotic treatment
aggravated by miss-diagnoses
UNIVERSITY OF SOUTH CAROLINA
Department of Computer Science and Engineering
Research Objectives
Build a Decision Support System to provide:
• Likelihood of a bacterial infection
• Measure of its severity
• Most likely site of infection
• Most likely pathogen
• Susceptibility of pathogen to drugs
• Gain in life expectancy through treatment
• Cost of drug treatment (price, side-effect, ecological
impact, future resistance)
• Ranking of anti-biotic drugs (Cost-Benefit)
UNIVERSITY OF SOUTH CAROLINA
Department of Computer Science and Engineering
Problems with initial BN Approach
• Model is not portable (specific to region/hospital)
– Dichotomous data (local vs universal)
• Human input error (20% of cases)
Obviating Enhancements
• Fix by normalizing system (localizing model)
• Fix by objective input requirements (symptoms, test results,
etc)
UNIVERSITY OF SOUTH CAROLINA
Department of Computer Science and Engineering
General Modularized Design
Nodes in BN:
•
Pathogen - Represent the potential pathogens of infections at the given site
•
M_Distrib – Major patient-groups exhibiting a particular pathogen within the given site
•
Minor – minor distribution factors; factors that change the likelihood of one or more pathogens without
affecting the overall risk for infection.
•
Infection – the different patterns in which infections manifest
•
Local-respo – local responses caused by an infection
•
Local-sign – manifestation of local responses.
•
Sys-respo – systemic response caused by infection and common to all sites of infections
•
Sign – manifestations of sys-respo
•
Spec-cultu – ability of pathogen to grow in local specimen
•
Blood-cultu – ability of pathogen to grow in the blood
•
Lab-site – ability of pathogen to grow at local site
•
Antibiotic_tr – antibiotic treatment prescribed for an infection
•
Coverage – the percentage of pathogens of a given infection susceptible to an antibiotic drug
•
Resistance – in-vitro susceptibility of pathogen to treatment
•
Cost – accounts for cost of using antibiotic: cost of purchase, side effects, and ecological impact
•
Gain – net gain in life expectancy gotten by prescribing an antibiotic drug
•
Underlying – disorders of the patient
UNIVERSITY OF SOUTH CAROLINA
Department of Computer Science and Engineering
General Scheme for Site of
Infection Network
UNIVERSITY OF SOUTH CAROLINA
Department of Computer Science and Engineering
Urinary Tract Infections Network
UNIVERSITY OF SOUTH CAROLINA
Department of Computer Science and Engineering
Calibrating the System
UNIVERSITY OF SOUTH CAROLINA
Department of Computer Science and Engineering
Results
• Addresses all important decision-points in first
days of patient care
• Expect the system to perform better than clinician
• No test data showing it improves clinical practice
and patient outcome - need a clinical trial
• Convenience of calibrating system for new
locations?
UNIVERSITY OF SOUTH CAROLINA
Department of Computer Science and Engineering