In a recent review published in the journal Cancer, research review publications investigating the benefits and limitations of AI precision medicine techniques in oncology research and treatment.
Study: Uses and limitations of artificial intelligence for oncology. Image Credit: metamorworks/Shutterstock.com
Background
The study focussed on the diagnostic and prognostic utility of artificial intelligence (AI) algorithms and discussed the impacts of AI-based chatbots (generative AI) in promoting anti-cancer outcomes in the last few decades.
Finally, they touch upon the current challenges to widespread AI deployment and suggest regulatory implementations that may bolster the performance of these algorithms in the coming years.
Precision medicine and its application in clinical anti-cancer applications
More commonly known as ‘personalized medicine’, precision medicine is the therapeutic approach that considers a patient’s specific genetic makeup, environmental exposure, and health behaviors (lifestyle and associated behaviors).
In contrast to traditional medical approaches, which primarily subscribe to the ‘one size fits all’ ideology, precision medicine presents numerous benefits, especially in the case of fields such as oncology, wherein patient-specific details (such as tumor information) can substantially improve clinical outcomes over general chemotherapy.
Innovations in oncology are of particular scientific interest, with reports revealing that cancer mortality rates have declined by more than 33% in the last 32 years alone.
Unfortunately, increased environmental pollution and suboptimal lifestyle choices have concurrently hampered progress in the field due to the increasing variability of carcinogens that trigger the condition.
Precision medical approaches, especially those that employ artificial intelligence (AI) algorithms, have the potential to overcome this limitation of conventional generalized medicine by allowing researchers and clinicians to identify better previously unknown patterns in patients’ radiological scans revealed by machine learning (ML) and deep learning (DL) technologies.
“AI algorithms are grouped into two categories: predictive AI and generative AI. Predictive AI tools learn patterns from training data to forecast outcomes in new scenarios. For example, an image-based classification tool used to diagnose breast cancer from mammogram scans is a predictive tool. Generative AI creates novel outputs that were not explicitly in the training data. AI chatbots that interact with patients in conversation are a form of generative AI.”
Unfortunately, despite developing and testing several AI algorithms for cancer care management, implementations of these technologies in mainstream medicine remain rare.
Notable roadblocks in incorporating AI models in the research included their relatively high upfront implementation costs, human noninterpretability of the algorithm outcomes, and limited human monitoring and validation of algorithms post-deployment.
Furthermore, research efforts in various aspects (and during different phases) of cancer care are not uniform, with substantially greater literature available on cancer diagnosis (>80%) compared to treatment and post-chemotherapy care.
Challenges notwithstanding, AI’s implementation in oncology has rapidly progressed the field, allowing for novel diagnostic, prognostic, and chat-based information access for both clinicians and their patients.
The present review discusses this progress and highlights the pros and cons of current AI implementations. It further discusses conventional and future challenges in widespread AI adoption.
It suggests policy changes that may further reduce the global burden of cancer, one of the most deadly and debilitating chronic diseases in the world.
About the review
The present review aims to provide context for three common use cases of precision medicine (particularly AI implementations) in cancer care – 1. Cancer classification and diagnosis, 2. Cancer prognostication, and 3.
Utility of AI chatbots and other large language model (LLM) technologies in optimizing clinical workflows.
It discusses the outcomes of more than 40 bodies of research (primary studies) to elucidate policy and implementation improvements that could further bolster cancer mortality rate reductions in the coming years.
Diagnosis
Cancer diagnosis, especially in early-stage cancers and cancers that have relapsed following previous treatment, given that most patients at these stages appear clinically healthy to human observers.
AI algorithms, especially ML ones, trained on millions of cancer diagnostic images (radiology scans, pathology images, and even patient-provided smartphone photographs) are efficient in identifying, classifying, and diagnosing these cancers, especially in cases where image data features are too subtle to be perceived by the human eye.
Even in cases where a diagnosis is a human preview, AI technologies, including computer-aided detection (CAD) algorithms (variants of DL frameworks), can highlight regions of interest (suspicious pixels in cancer diagnostic images) to aid clinicians in their diagnosis evaluations.
Surprisingly, AI algorithms have, in some cases, displayed better diagnostic accuracy and efficiency than their human counterparts.
“Commonly used AI algorithms for image classification are convolutional neural networks (CNN), deep learning architectures that extract identifying features for each group and use the resulting schema for a new classification task. The algorithm assigns a probability for each output class, and the image is classified into the group assigned the highest probability. The accuracy of the AI tool is measured by comparing the algorithm classifications with clinician classifications, referred to as “ground truth”.”
The major pro of AI implementation in diagnosis is melanoma and breast cancer screening, where early detection is the most important variable in favorable mortality and morbidity outcomes. Unfortunately, AI suffers from severe training-associated biases, significantly hampering its implementation in the field.
Underreporting of training data, alongside inconsistent representation and data heterogeneity (image acquisition and processing), makes most AI models non-generalizable, preventing their incorporation into global oncology protocols.
“Modifications along the algorithm development pipeline can help mitigate these concerns. Training data can be expanded to include representative images from all demographics (e.g. skin color, ages, and body types). Training sets with image data should include samples taken from different angles, lighting, and equipment; and AI technologies should accommodate changes in image acquisition technology by retraining the model with new images.”
Prognostication
Forecasting patient outcomes is one of the most essential early-stage clinical intervention steps carried out by medical practitioners, as it allows clinical interventions to be tailored to improve or avoid the most adverse clinical outcomes.
Unfortunately, human-conducted prognostication is historically susceptible to significant error, with reports estimating that 63% of prognoses are overestimations of outcomes, while 17% underestimate patient survival.
“The consequences of inaccurate predictions in oncology include increased emotional burden on patients and their caregivers, inappropriate allocation of resources, decreased trust in the patient–physician relationship, and delay in crucial therapeutic or end-of-life interventions. AI-based risk prediction models that generate individualized estimates on prognosis have augmented clinician assessments of risk and aided personalized care decisions in oncology.”
Electronic health records (EHR)-based ML models have shown great promise in this field. They have been proven to predict cancer outcomes months or even years in advance, thereby allowing clinicians the information they need to best prepare for the oncological eventuality.
Moreover, these models can evaluate the most efficient and cost-effective clinical intervention route, thereby saving extensive (clinical) human resources and (patients’) financial investment, reducing the overall disease and socioeconomic burden of the disease.
Unfortunately, most of these models are deterministic in nature and are thus susceptible to changes in model results on the inclusion of novel, yet computationally unaccounted for, data generation approaches.
‘Performance drift,’ the gradual decline in model performance over time, can make subsequent model predictions inaccurate and unreliable unless frequent updates to its modeling algorithm and human results validation are routinely carried out.
In this field, quality of training data, frequent human model validation, and data-sharing across different cancer types may overcome these challenges in the future.
Chatbots and conclusions
Modern conversational chatbots, particularly platforms such as ChatGPT, Google Gemini, Microsoft Copilot, and others, are revolutionizing the way both professionals and laypeople acquire and process information from the World Wide Web.
These generative AI applications are designed to harness the power of LLMs to output novel content in based on the user’s need.
Unfortunately, research into the applications of chatbots in oncology has revealed that the technology is still in its nascent stages with little to no support, let alone policy-approved implementation in clinical practice.
“The adoption of chatbots for medicine relies on achieving both understandable language and conveying complex medical topics accurately, which current algorithms cannot do consistently because readability scores vary by the user’s verbiage of the prompt. Although medical knowledge expands each day, algorithms are not continuously updated to accommodate this change. As a result, the chatbots that are not trained on updated information can become unreliable and more inaccurate with time.”
Together, these individual, field-specific pros and cons paint an interesting picture – while the importance and relevance of AI implementation in oncology research cannot be overstated, these models’ computational and raw data requirements are only recently beginning to be met.
With the development of improved modeling frameworks, the Availity of larger and higher-resolution datasets, and increased scientific verification of their accuracy and reliability, AI models present a powerful tool in the oncologist’s arsenal against this terrible disease and may one day take a majority of the cancer care burden off human medical practitioners.
