1155P - Application of novel machine learning to predict immunotherapy related toxicities for metastatic melanoma patients from baseline 18F-FDG PET/CT scans

Background

Immunotherapy (I/O) has been shown to provide durable responses in metastatic melanoma (MM) patients, however, I/O related adverse events (irAE) are frequently experienced and can be challenging. ¹⁸F-FDG PET/CT medical imaging, has the potential to predict both tumor response and irAE non-invasively. Developing machine learning (ML) models to quantify features derived from medical images represents an opportunity to predict toxicity and influence clinical patient care. This study implements automated organ segmentation and ML methods using ¹⁸F-FDG PET/CT images to predict irAE for patients with MM.

Methods

¹⁸F-FDG PET/CT scans from 128 patients with MM were retrospectively collected between 2009 and 2021 (IRB approved protocol). Patients received ≥ 1 course of I/O. Organs were segmented automatically using AIQ Solutions technology. Segmented regions were used to quantify FDG organ uptake and correlate with toxicities including adrenal insufficiency, alanine aminotransferase increase, colitis/diarrhea, pancreatitis, and thyroid dysfunction. Imaging features were extracted from organs at baseline (BL) and early follow-up. Organ uptake and changes across time were evaluated for predicting irAE. A random forest model was trained using BL organ uptake to predict occurrence of an irAE. Performance was evaluated using area under the receiver operating characteristic curve (AUC).

Results

The patients (89 male and 39 female) average age was 63 (range 23-88). Overall, there were 105 irAE grouped per organ for univariate analysis, with the high frequency of hypothyroidism, raised alanine aminotransferase and colitis/diarrhea. The overall AUC for prediction of any irAE based on the BL scan was 0.76. Change in 95^th percentile between pre-I/O work up and subsequent scans of SUVs was predictive of irAE in thyroid (AUC=0.95) and bowel (AUC=0.94).

Conclusions

Our results indicate that ML using quantitative features from ¹⁸F-FDG PET/CT imaging has the potential to identify irAE early. Importantly, there may be imaging features on the BL scan that indicate MM patient’s likelihood of developing an irAE before starting I/O. This warrants further investigation in a prospective setting.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Has not received any funding.

Disclosure

All authors have declared no conflicts of interest.