Research Article By
From ancient Egyptian mummies, we know that cancer has been a part of civilization since the earliest times of recorded history. Cancer is fundamentally a disease of gene expression and regulation. Research has focused on the pathways around promoting cell division (e.g. Ras and the promotion of excessive cell division) or the pathways around inhibiting cell growth (e.g. p53 where the inhibition of a transcription factor results in a lack of cell division. Multiple gene mutations are typically required before the development of cancer. The combination of multiple locations as well as multiple pathways are all opportunities for cancer detection and treatment.
A biomarker is any objective measure which can be used to describe current clinical state. Examples include blood pressure, the presence or lack of presence of a gene, the amount of a chemical quantity. For cancer, example biomarkers measured via, say, blood, are calcitonin (thyroid cancer), CA-125 (ovarian cancer), or AFP (liver cancer). Combined with imaging approaches such as CT scan, MRI, or X-rays, even amongst the very short list we have mentioned, a fundamental question arises: which of these biomarkers, and in what combination constitutes a diagnosis for cancer? Or even better, precancerous conditions? Machine Learning and AI represent one of the new frontiers in cancer detection using biomarkers as the training inputs into predictive models for cancer detection and treatment.
It is a combination of factors that paints a picture of whether a set of biomarkers constitute a diagnosis for cancer, particularly when the goal is early detection where the traditional biomarkers for a particular type of cancer may be inconclusive. AI helps as it can uncover relationships that are hard to detect amongst large sets of seemingly unrelated factors. AI requires a large amount of data, then algorithms to process the data as a human brain would in order to look for patterns and variations. AI is being used in healthcare to analyze data and assist in diagnosis and treatments. For example, it can analyze data from a large population of patients and predict health risks and outcomes. In order to perform such tasks, the systems are trained by humans; algorithms are written to project how a human would handle the situation whether it be finding patterns in data or diagnosing patients. AI allows the healthcare industry to be representative and predictive to each patient.
The increasing rate of publications about AI in biomedicine. (source)
Artificial intelligence consists of software frameworks known as artificial neural networks, which work like a human’s interconnected neurons in the brain, resulting in outputs to respond to stimuli. The neural networks are trained by humans through three different types of learning. Supervised learning is when the neural network has to predict input and output pairs based on example pairs,
which are known as the training data. Unsupervised learning is when the neural network has to learn from an unclassified dataset to find repeating features across the data. The Principle Component Analysis and clustering methods aid in the feature identification unsupervised learning is responsible for. Lastly, reinforced learning is when the neural network must get the most rewards possible and the least consequences possible within a given area. Deep learning is a newer form of neural network methods, which uses neural networks with many more layers, allowing the software to mimic the human brain even more. Apart from machine learning, AI’s function also consists of natural language processing which allows for information extraction from unstructured data such as notes and journals, that do not always have a specific format they are written in.
Data modalities and opportunities for multimodal biomedical AI. (source)
Artificial intelligence is mainly used to aid in clinical decisions and medical information management in the healthcare industry. For example, Apple uses AI in healthcare by incorporating a sensor in their smartwatch, the Apple Watch, that detects Atrial Fibrillation (AF). The watch does this by detecting irregular pulse rhythm because AF is often asymptomatic or associated with the risk of stroke. In fact, 34% of the Apple Watch users that received an AF detection notification tested positive for it. 3
AI has other uses in health care such as the detection of dementia. Currently, even without AI, we are able to detect dementia, but we are forced to use really invasive procedures or use really expensive methods, which are barriers preventing widespread use. At the University of Cambridge, a team led by Professor Zoe Kourtzi created tools that used machine learning and AI to detect whether or not a patient had dementia from a single brain scan, even before any sign of symptoms. Their algorithm spotted any structural changes in the brain. This system had an 80% accuracy rate. Another way AI is helpful is a database: conventionally, patients who were suspected with dementia would need to take a paper and pencil test, but by entering all the patient’s information and test results into a database, the doctor is able to get guidance for next steps from the system. The database is able to compare the information of the current patient with other patients predicting on whether the patient has dementia or not.
The process where unstructured data is read by the machine’s software and then goes through the machine learning processes to be analyzed furthermore. (source)
AI for screening of multiple retinal and optic nerve diseases are being used in a number of ways including, through the use of RAIDS and through the use of AI modeling. RAIDS, otherwise known as Random Active Image databases are used in AI screening for retinal and optic nerve diseases to store different types of scanned images to identify AI algorithms for the screening of retinal and optic nerve diseases such as diabetic retinopathy, age-related macular degeneration, glaucoma and optic neuritis. RAIDs also help ensure that AI algorithms can accurately recognize the disease in a diverse patient population. The use of RAIDs allows for the integration of large amounts of diverse data and the ability to create large datasets for the training and validation of AI algorithms.
AI and radiomics are used in many ways for lung cancer screening. For example, radiomics are used in lung cancer screening to better improve the accuracy and efficiency of screen testing for early detection of lung cancer.** Computer-aided lung nodule detection is a type of AI-powered image analysis technology that can detect small lung nodules. These nodules are highlighted and evaluated by radiologists to see if they are linked to lung cancer. The AI analysis technology also helps with reducing the burden on radiologists by reducing the time and effort required to manually review images and identify lung nodules. Deep learning is another type of machine learning that uses algorithms that analyze medical images, like CT scans, to detect lung or pulmonary nodules too. All these AI analysis programs help detect these nodules and signal radiologists about their link to lung cancer.
The application of AI to cancer detection is ever evolving, and especially as we better understand what factors are highly correlated with identifying cancer, AI models will only get better at their predictions and become more accurate, but they will also identify relationships we did not understand until in hindsight they were identified as correlated.As more data is collected and algorithms are developed, the increased use of AI for consultations, diagnosis, and educational purposes. For future patients, doctors with the ability to early detect cancer represents progress in the ability to treat it.
Peter Chen (High School)
Romtin Pourzand (High School)
Sriya Pasala (High School)
(External students, post on the request of Mentors)
The opinions expressed here are the views of the writer and do not necessarily reflect the views and opinions of Elio Academy.