SVS presentation 5/24

BL
by Becky Long0 views·Updated May 25, 2022
Re-use contentCreate your own

Chapters

Audio TranscriptAuto-generated

  1. Good morning.
  2. Thank you for the opportunity to present our research
  3. on applying artificial intelligence to predict complications after Eva.
  4. Although the endovascular treatment of triple A.
  5. Has become the standard of care complications such as Angelique
  6. graph migration and rupture are associated with high morbidity and mortality
  7. patients at high risk for such events should have more frequent surveillance.
  8. Although this can be difficult due to lack of patient compliance,
  9. artificial intelligence has revolutionized many areas of research.
  10. Prior efforts in vascular surgery to predict
  11. evil complications were hindered by imbalanced,
  12. low complication rate and skilled data distribution.
  13. Our aim was to develop an Ai model
  14. to predict post evil complications for individual patients.
  15. We review EMR and operative logs to identify all patients. From 2010 to 2020.
  16. Pre op C. T. A. And three D reconstruction images were obtained
  17. 273 patients were included. Average age was 74 years old, 81% were male.
  18. Average triple a. size were 5.3 cm.
  19. The C. T. A.
  20. Exams had been performed according to the hospital's usual protocol,
  21. actual images of education were used by the pACS
  22. software to generate the three D reconstruction images.
  23. These unmeasured images from the pACS system were exported into a J. P. E. G. Format.
  24. After the identification they were uploaded to google
  25. collab as the input for model development.
  26. This is a picture of the neural network.
  27. The model backbone utilized convolutional neural network of multiple layers
  28. wherein imaging are independently processed
  29. for object detection and recognition.
  30. With mathematical model
  31. simulating some of the property of the visual cortex.
  32. Training,
  33. the model involved two important steps because complications are infrequent,
  34. there's an imbalance in the data.
  35. Very negative examples dominate the sample population.
  36. Therefore down sampling was randomly applied so that
  37. the positive and negative classes had similar numbers.
  38. While down sampling can alleviate the imbalance issue,
  39. it might also negatively impact the model sensitivity and
  40. negative predictive volume due to the loss of data.
  41. To compensate for this,
  42. a data augmentation technique was employed to re enlarge the dataset size.
  43. A total of 40 positive 189 negative cases
  44. were randomly selected to form the training dataset
  45. in data augmentation.
  46. An original imaging in the center was used as a template from which
  47. um New images were deployed, developed through software by rotation, cropping,
  48. discoloration,
  49. shrinking and blurring. These additional images were uploaded into the model.
  50. While the processed images look similar to the original ones,
  51. they were considered different.
  52. New data to the model. This enlarged the dataset fourfold
  53. For the testing phase.
  54. The remaining a positive and 36 negative cases were
  55. used to test the performance of the model.
  56. The training and testing datasets were completely
  57. independent of each other and only images that
  58. had not been seen in the training process were used in the testing phase.
  59. A total of 17 patients develop complications of
  60. which the majority were end Alex type two
  61. were the most common type of Angelique.
  62. I mean follow up period was 46 months and five post op visits.
  63. There are no differences in clinical characteristics
  64. between the patients with and without complications.
  65. Overall the model achieved the desired
  66. high sensitivity and negative predictive volume of
  67. 100% and identified all the patients who later developed complications after Eva.
  68. The clinical priority is not to miss any life threatening complications.
  69. Even though this increased the rate of
  70. false positive and decreased the overall accuracy.
  71. As a practical clinical application.
  72. Such results would lead to identification of patients at high
  73. risk and more robust efforts at recommended patient surveillance.
  74. As a more detailed evaluation of the model the our C. Curve and A U. C.
  75. Score of 0.62 in the testing phase
  76. illustrates very good performance in predicting patient complications.
  77. In conclusion,
  78. modern AI model offered the potential
  79. to effectively predict complications after Eva.
  80. With down sampling and augmentation technique.
  81. The West Ai model developed in this study suggests 100% sensitive sensitivity
  82. um which can help to guide and optimize patient post up surveillance.
  83. Hopefully future directions will include
  84. external validation and model improvement
  85. within large data sets in collaboration with other centers.
  86. The incorporation of inter and post operative images into the model and deep
  87. dives into the neural network layers to
  88. identify and quantify predictors and specific complication
  89. preet abilities.
  90. Thank you for your attention