Leveraging AI for More Accurate Multiple Myeloma Predictions
Artificial intelligence shows promise in improving diagnosis and treatment for patients with multiple myeloma.
The interest in and use of artificial intelligence (AI) has grown substantially in recent years across industries, and its capabilities show great promise for advancing modern medicine through advanced screening, diagnostics, and treatment of diseases. At the International Myeloma Society 21st Annual Meeting in Rio De Janeiro, Brazil, Anant Madabhushi, PhD, professor of biomedical engineering at Emory University and the Georgia Institute of Technology, discussed the recent findings and opportunities for AI to create better predictors for diagnosis, prognosis, and treatment response for patients with multiple myeloma (MM).1
The use of AI in the treatment of patients with MM is in its infancy, and continued investigations are needed. Image Credit: © AIproduction - stock.adobe.com
Madabhushi opens his presentation with a discussion on the issue of overdiagnosis and overtreatment of disease, citing the subsequent financial toxicities and significant harm caused by aggressive treatments. Through the example of prostate cancer, he suggests that indolent diseases with lower mortality rates and higher survival outcomes are often treated heavily with chemotherapy, radiation, and surgeries with risk of adverse effects and recovery burdens. Additionally, there are the unavoidable financial impacts on patients. According to the study “Death or Debt? National Estimates of Financial Toxicity in Persons with Newly-Diagnosed Cancer” published in the American Journal of Medicine, a little over 42% of patients with cancer deplete their life savings within 2 years of their diagnosis.2
Through use of AI, clinicians can better extract various tumor characteristics, such as proteomic and genomic data, as well as mappings of physical structures and tumor architecture to more accurately characterize disease. Machine learning models can then be employed to combine all of this information to create better predictors for diagnosis, prognosis and therapeutic response in patients. Through use of AI tools, clinicians may be able to circumvent more invasive methods and potentially replace unnecessary treatments with more targeted, precise therapies.
“Initially, when we did this work, we were really taken aback by how powerful these algorithms could be,” said Madabhushi. “We were able to stand up an AI algorithm to now go and find cancer cells on new, unseen images. But one of the challenges with these deep learning or AI algorithms is that they're often referred to as ‘black box algorithms,’ and the reason for that is because we don't truly understand what the AI is learning.”
A significant obstacle, called the “black box algorithm,” is the lack of interpretability can lead to a reproducibility crisis in clinical science, making it challenging to understand how models will generalize to new data across different institutions. Additionally, there is the risk of these models “hallucinating” or producing convincing, albeit incorrect outputs. This highlights the need for continued refinement of models and algorithms, as well as the integral role of the human mind in overcoming these obstacles. There is a growing need for more explainable AI algorithms capable of analyzing a patient’s tumor characteristics and providing accurate data to inform more personalized treatment curation. One way to make this happen is through the inclusion of AI tools in blinded clinical trials to produce high evidence generation.
Leveraging AI for the deep analysis of tumor and malignant cell characteristics may not only be useful for diagnosis and progression, but also the prediction of treatment response. In a study of patients with early-stage breast cancer, analysis of routine breast pathology images allowed researchers to differentiate low-risk from high-risk patients using the patterns and orientations of collagen fiber in the tumor microenvironment. Patients with low collagen fiber entropy had worse outcomes and those with high entropy had better outcomes. Researchers found that 12% of patients classified as low risk through molecular assay were high-risk based on the collagen fiber patterns, suggesting they may have benefited from chemotherapy that they did not receive. Further, they observed that 57% of patients who were classified as high-risk by the molecular assay were identified as low risk by the collagen fiber analysis, indicating they may have been over-treated with chemotherapy.
“What we're showing here is that more explainable, interpretable features could add significant value to the status quo, and allow for more granular risk stratification compared to what might be obtainable from more expensive tissue destructive molecular-based assays,” said Madabhushi.
He highlights AI’s use for other disease states, including multiple myeloma (MM). By analyzing fundus images from the UK Biobank, Madabhushi and his team observed that tortuosity of the blood vessels in the eye appeared to predict the onset of MM within 10 years, suggesting the potential use of routine eye imaging and AI to identify at risk patients. There are additional efforts to use machine learning for rapid prediction of M-spike values, an important biomarker for monitoring MM.
The use of AI in the treatment of patients with MM is in its infancy, and continued investigations are needed to determine whether its capable of providing accurate, actionable predictive insights. As AI continues to evolve in MM care, it's critical that these advancements deliver precise insights and also address disparities in access and equity.
“We have to think about low-cost tools that can really have impact in low-middle income countries, and we have to think about equity,” concluded Madabhushi. “We have to make sure that we are very intentional as we're developing these AI tools to make sure that they work across diverse populations.”
