Recurrent Recognition Networks for Approximation of BN2O Inference Dedicated to the Medical IntMed System
Author
Henriksen, Jonas Tvede
Term
4. Term
Publication year
2021
Submitted on
2021-06-11
Pages
87
Abstract
Dette projekt undersøger en alternativ metode til inferens i en stor grafisk BN2O-model i AI-systemet IntMed, som bruges til klinisk beslutningsstøtte under realtidskonsultationer. Ambolt ApS har leveret de data, der gør undersøgelsen mulig. Den nuværende algoritme, Quickscore, beregner eksakte sandsynligheder, men har svært ved at overholde realtidskravet. Vi foreslår at bruge såkaldte Recognition Networks (neuronale netværk, der lærer at tilnærme opdaterede sandsynligheder) til approksimativ inferens. Konkret præsenterer vi et rekurrent genkendelsesnetværk, som kan udnytte den tidslige rækkefølge, hvori symptomer kommer frem under patientudspørgning. Vi viser, hvordan dette rekurrente netværk kan trænes med forward sampling (fremadrettet sampling) fra BN2O-modellen, dvs. ved at generere simulerede patientforløb. For at demonstrere potentialet sammenligner vi netværket med Quickscore i forskellige konsultationsscenarier med fokus på posterior-sandsynligheder (opdaterede sandsynligheder) og rangordning af sandsynlige sygdomme. Resultaterne viser, at et rekurrent neuralt netværk bedre kan efterligne eksakt inferens end en ikke-rekurrent, sekventiel modpart, men at der fortsat er behov for opmærksomhed på skalerbarhed og kalibrering.
This project investigates an alternative way to perform inference in a large graphical BN2O model within the AI system IntMed, which is used for clinical decision support during real-time consultations. Ambolt ApS provided the data for this study. The current algorithm, Quickscore, computes exact probabilities but struggles to meet strict real-time demands. We propose using Recognition Networks (neural networks trained to approximate updated probabilities) for approximate inference. Specifically, we introduce a Recurrent Recognition Network that can exploit the order and timing in which symptoms are revealed during patient questioning. We show how to train this recurrent network with forward sampling from the BN2O, i.e., by generating simulated patient cases. To demonstrate its potential, we compare it with Quickscore across various consultation scenarios, evaluating posterior probability estimates and the ranking of likely diseases. Our results indicate that a recurrent neural network mimics exact inference more closely than a non-recurrent, sequential counterpart, but still requires attention to scaling and calibration.
[This summary has been rewritten with the help of AI based on the project's original abstract]
Documents