Tag Archives: machine learning

Will Future Pathology Reports Include Likely Therapeutic Recommendations?

Posted on by

X Liu et al. Gastroenterol 2024; 166:921-924. Machine Learning–Based Prediction of Pediatric Ulcerative Colitis Treatment Response Using Diagnostic Histopathology

M Iannucci et al. Gastroenterology 2024: 166: 730-732. Editorial. Open Access! A Baby Step or a Real Giant Stride: Histomic Enabled by Artificial Intelligence to Predict Treatment Response in Pediatric Patients With Ulcerative Colitis

In this article, their machine learning algorithm was trained on 187,571 informative patches from rectal hematoxylin and eosin biopsy samples from 292 treatment-naive pediatric patients with UC in a multicenter inception cohort (PROTECT) 2) study.

Key findings (summarized by editorial):

  • The authors first trained the machine learning models on 250 histomic features at the patch level and achieved an area under the receiver operating characteristic curve (AUROC) of 0.87 (95% confidence interval [CI], 0.73–1.00) and an accuracy of 0.90 (95% CI, 0.80–1.00) at the WSI (whole slide images) level in predicting treatment response.
  • A subset of 18 histomic features exhibited comparable performance with an AUROC of 0.89 (95% CI, 0.71–0.96) and accuracy of 0.90 (95% CI, 0.80–1.00) to model using the full set of 250 features, indicating the potential for standardized practical application in clinical settings.
  • The authors confirmed that the set of 18 histomic features demonstrated comparable performance on the real-world independent SickKids cohort7 (University of Toronto) with an AUROC of 0.85 (95% CI, 0.75–0.95) and accuracy of 0.85 (95% CI, 0.75–0.95) at the WSI level.

An important limitation on this study was that the population was 83% white, indicating that it may have less applicability in other cohorts.

My take (borrowed from editorial): Although this study focused solely on the therapeutic outcomes of mesalamine on treatment-naïve patients, it is anticipated that a comparable methodology (based on the fusion of machine learning and digital histopathology) could be applied in subsequent research to elucidate the necessity for colectomy, evaluate responses to biological agents, and optimize drug selection from the current armamentarium. In other words, this approach utilizing routine biopsy specimens may offer a pathway toward a precision personalized approach to managing pediatric-onset UC (and possibly IBD more broadly).

Torrey, Utah

#NASPGHAN19 Impact of New Technologies on Patient Health

Posted on by

Along with Ragh Varier, I had the privilege of moderating a session on new technologies on patient health.  Below I’ve included a few slides and some notes; my notes may have errors of omission or transcription.

Chicago

 

Dr. Mehta’s lecture focused on wearable health technologies. Key points:

  • It is already in use in some areas (eg. continuous glucose monitoring for diabetes, ECG sensors).
  • She noted that wearable technology dates back to the 1600s with the abacus ring
  • Challenges: Accuracy, Actionability/outcome improvement, Reaching at-risk populations (not just the ‘worried well’ populations), regulation, sustainability (users may abandon quickly), and ethical/privacy concerns
  • Some families taking technology into their own hands, so to speak. #WeAreNotWaiting.  Example: artificial pancreas device system

Dr. Syed’s lecture focused on artificial intelligence in medical-decision making. Key points:

  • AI is already in use in areas like facial recognition
  • AI may be able to increase polyp detection rate in colonoscopy and improve histology reading
  • Her team has been working on using AI to help distinguishing enviromental enteropathy histology from other etiologies
  • Other potential uses: AI to help predict Crohn’s disease progression based on histology

Related study (not discussed in talk): Z Deng, H Shi et al. Gastroenterology 2019; 157: 1044-54. The authors collected more than 113 million images from 6970.  With a deep-learning algorithm, they found that video capsule endoscopy could have higher detection rates and improved reading time with a “CNN-based” reading system (CNN=convolutional neural network).  The mean reading time was reduced from 97 minutes with conventional reading to 6 minutes with CNN-based reading system.  The later had 99.88% sensitivity in per-patient analysis (vs. 74.57% with conventional reading).

The oral abstract presentation, by Sonja Swenson, detailed how machine learning was applied to try to improve transplantation selection/PELD scores.

  • The authors of this abstract (437) used data from 6273 patients with PELD scores and added additional variables to try to identify a more accurate model.
  • Link: All NASPGHAN 2019 Abstracts

Dr. Li, known by some as the ’emperor of emesis,’ presented a lecture on telemedicine. His full slides: Telemedicine NASPGHAN Updated 2019 (B Li)

Key points:

  • When surveyed, patients/families prefer telemedicine over conventional medicine.  Key reason is convenience
  • Lots of issues from health care provider viewpoint: reimbursement, licensing (improving), increased time
  • Many examples of telemedicine/telemonitoring that are ongoing

Disclaimer: NASPGHAN/gutsandgrowth assumes no responsibility for any use or operation of any method, product, instruction, concept or idea contained in the material herein or for any injury or damage to persons or property (whether products liability, negligence or otherwise) resulting from such use or operation. The discussion, views, and recommendations as to medical procedures, choice of drugs and drug dosages herein are the sole responsibility of the authors. Because of rapid advances in the medical sciences, the Society cautions that independent verification should be made of diagnosis and drug dosages. The reader is solely responsible for the conduct of any suggested test or procedure. Some of the slides reproduced in this syllabus contain animation in the power point version. This cannot be seen in the printed version.