The virtual AACR conference featured a panel discussion on the development of Artificial Intelligence (AI) and Machine Learning (ML) in cancer research, where speakers addressed the challenges and progress in the area. They also acknowledged the structural inequities in the healthcare system that needs to be accounted for while training algorithms for clinical care and diagnosis.
Dr. Iya Khalil from Novartis talked about multiple sources and levels of patient data that can be integrated with clinical trial data to bridge the gaps and provide a wider representation of the real-world manifestation of any disease. In her opinion, the responsibility of data science lies in understanding methods and mechanisms underlying disease from multiple streams of data. She argued for the integration of multiple modalities of data such as molecular data, images, biopsies, etc which can help remove some bias.
Dr. Iya Khalil, Novartis Institutes for Biomedical Research
Dr. Leslie from the Memorial Sloan Kettering Cancer Center talked about the risks of training an algorithm with data from a small and uniform population and how that can create unstable models which could miss important data points. In basic research, the end goal of AI/ML is accurate predictions based on input data. However, in clinical practice, AI/ML needs to be able to accurately predict and also possibly explain and interpret the mechanism which enables that prediction for a clinician to be able to trust it.
Dr. Elmore from the Fielding School of Public Health addressed the difficulties of dealing with the real-world implications of false-positives in AI/ML guided diagnosis. Dr. Lin talked about the differences in clinical trials in controlled environments and real-world clinical care which can be minimized by AI/ML.
Discussing if technology will exacerbate or reduce the digital divide among over- and under-represented populations, each panelist underscored the need to develop tools and regulatory oversight which can make the playing field equal across all healthcare centers to have the same level of clinical care, while being aware that input data has some inherent bias.
In the accompanying pre-recorded sessions, the panelists shared their experiences of using AI/ML for cancer research and care.
Using Big Data to Study Gene Regulation
Dr. Christina Leslie talked about leveraging big data from epigenetic studies to decode chromatin codes and gene regulation. Dr. Leslie’s lab focuses on combining DNA sequence, information about accessibility and activity of regulatory elements such as promoters and enhancers by Hi-C and HiChiP data sets from various physiological states into an AI/ML module to predict gene expression.
Dr. Christina Leslie, Memorial Sloan Kettering Cancer Center
They used interactive 3D genome data to develop a graph neural network- GraphReg, which can identify highly significant interactions between regulatory elements by novel statistical analyses. They trained GraphReg to predict long-range interactions in the genome for better gene expression modeling and prediction of functional enhancers by using feature attribution. Trained on cell line data, GraphReg performs well with a good correlation between true signal and predicted signals and better than baseline sequence models which only use 1D data.
They designed two versions of GraphReg:
- Epi-GraphReg- It uses 1D input and histone modifications to generate conventional neural networks combined with graph data from HiChiP to predict gene expression from activity and connectivity of promoters and enhancers. It is cell-type agnostic.
- Seq-GrapgReg- input is `DNA sequence into cnn and graph attention network predicts gene expression and epigenetic signals. It is cell-type-specific.
To understand how predictions are made, Dr. Leslie’s team looked at features from the ML model which enables it to interpret the data into significant regions and contacts to predict gene expression. Dr. Leslie concluded that a combination of ML, epigenomics, 3D genomics, and gene editing have contributed to improved modeling of gene regulation.
Translating AI/ML Innovation into Practice
Dr. Lynda Chin of Apricity Health shared her views on the necessity to create an ecosystem to translate AI/ML innovation into practice. Single cell genomics, DNA sequence, quantifiable behavior from wearable tech, social determinants can be combined to enable precision medicine at the individual level. In her view, precision medicine is most effective when it is data-driven, augmented by AI, and aided by technology. It shows promise in integrating clinical evidence, validation of evidence, and real time decisions for patient care.
Dr. Lynda Chin, Apricity Health
She addressed many inequities that need to be minimized to enable the effective translation of AI/ML-derived data into healthcare and provided an example of how innovation in AI/ML can help patient care. Big data will continue to increase in complexity and volume and hence we need smart approaches to continually narrow the gap between clinical research and patient care. Technology can help develop virtual points of care where they can receive health solutions at home. Patient data derived from multiple tools and analytics need to be contextualized by a clinician considering insights from their interaction with the patient to drive suitable clinical actions and ensure the best outcome. Tools can help clinical decisions but do not treat patients.
Apricity is a digital care platform with AI which leverages AI/ML to provide patient-centred cancer therapy at their homes. Immuno-Oncology (IO) therapy has been approved for a wide spectrum of cancers. However, IO is also associated with high toxicity and side effects in the real world causing a burden on the healthcare infrastructure and financial strain for patients.
With the Apricity app, patients can log in their symptoms and additional information during their IO therapy. The algorithm integrates the input into their medical records and sifts it for clinical relevance and need for clinical intervention, patient’s response to IO, to enable better treatment. The algorithm can trigger sample collection in response to an adverse event during the treatment which can capture real-time data of therapy and its side effects to inform subsequent steps in patient care and better clinical decisions.
Dr. Chin made a case for the next generation of IO therapy which combines AI-enabled real-time data collection, predictive analytics can be fed into a platform to learn disease progression and response of each patient. Precision oncology care can combine patient data, AI and digital technology to minimize the gaps between clinical trials and real-world outcomes and can make healthcare more equitable.
Clinician Variability in Cancer Diagnosis
Prof. Joann Elmore demonstrated clinician variability in cancer diagnosis. Emphasizing the importance of a correct diagnosis in cancer, Dr. Elmore argued that it drives treatment, surveillance, prevention, and screening. Interpretation of medical imaging is challenging since the same image can render multiple diagnoses, suggesting that there are no published guidelines and protocols for radiologists to make an accurate diagnosis.
Dr. Joann Elmore, UCLA Fielding School of Public Health
From personal experience, Dr. Elmore recounted obtaining 3 different diagnoses on the same skin biopsy. In an earlier study in 2015, she found that there was significant variability in diagnosis on the same breast cancer biopsies by many pathologists. Currently, she works on the Melanoma Pathology Study to quantify the accuracy and reproducibility of diagnosis of skin biopsies for melanoma.
Using a consensus diagnosis from 3 experts as a reference, she compared the diagnoses from over 250 pathologists for 240 cases to be classified into 5 classes with mild cases assigned to class 1 and invasive melanomas, class 5. Each participant made the diagnoses twice separated by a 6-month period. Strikingly, the participants matched the expert consensus diagnosis fairly well in mild/benign cases and invasive melanomas and were significantly variable in the middle. This presents a challenge in treatment, outcomes, and subsequent progression of the disease.
Since the participants made the diagnoses twice, intra-observer concordance was again high for classes 1 and 5 and low for intermediate classes. Some of the variability could be attributed to limitations of the test study and other factors such as the patient’s age, availability of healthcare options, and second opinions. Hence, Dr. Elmore talked about how AI needs to be trained to accommodate for these social determinants and clinical variability to help with an accurate diagnosis.
She talked about how computer-assisted detection (CAD) tools for biomedical image analysis can both help and hinder diagnosis. Hence, we need to exercise caution in developing AI/ML tools for biopsies. Some of the issues she pertinently raised were, bias in input data for training, what constitutes a consensus/reference diagnosis, how diagnosis informs clinicians’ interactions with their patients, and how CAD does not always necessarily improve diagnosis and may increase false positive rates and financial burden on patients.
©www.geneonline.news. All rights reserved. Contact: email@example.com