Abstract
The application of artificial intelligence (AI) to precision oncology has the potential to significantly enhance cancer treatment by personalizing care based on individual patient data. It can help identify potential drug targets, simulate drug interactions, personalize treatments, and potentially accelerate the development of new treatments. AI can analyze vast amounts of clinical data, including patient records, genetic information and imaging data and it can identify patterns, detect correlations, and make predictions that may aid in identifying potential treatment outcomes, adverse events, or drug interactions. 1 Predictive analytics enabled by AI can help anticipate how patients are likely to respond to different treatments. By analyzing historical data and patient-specific variables, AI can predict the efficacy of various therapies, potential adverse effects, and overall prognosis.
However, overcoming challenges related to risk averseness, data quality, collaboration, regulation, cost and skills development are needed to realize its full potential.
The highly regulated nature of the pharmaceutical sector and the inherent risk averseness of leadership creates a bias towards incremental improvement. This is manifest by the tendency to favor proof of concepts and pilots when it comes to new technology. However, many pilots and proof-of-concept studies—perhaps up to 90%—are not likely to move into production. The crowded vendor market also contributes to the bias towards pilots and proof of concept studies. For example, Forbes received over 1900 submissions in compiling the list of the top 50 AI companies. Not surprisingly, technology vendors compete for business and gladly accept small projects.
Data challenges drive the relative low success rate of pilots in the application of AI to precision oncology. Applying AI in precision oncology involves taking advantage of large volumes of data from diverse sources, such as genomics, radiomics, and clinical information. Data tends to be housed in disparate systems and on different platforms. Data quality issues in electronic health records due to duplicate entries and typing errors are common. Updating data in a standard format is challenging. Curating data repositories requires consistent attention on cycling out old information Genomic data and clinical data tend to come in various formats with little standardization, leading to further challenges and issues. Such data related concerns along with human biases that trickle into algorithms during development and post-deployment phases may affect performance thereby limiting the utility of AI technology in precision oncology. In turn, the potential for error may erode trust in new technology. In a nutshell, the success of AI models in precision oncology depends on high-quality, data in standard formats as AI algorithms are only as good as the data and assumptions that are used.
A related challenge to data issues is the lack of adequate and clear FDA regulatory guidelines along with high upfront costs for the integration of AI into clinical workflows, non-interpretability of the algorithms, and limited monitoring of algorithms post-deployment.
Broad-based collaboration is needed for success in applying AI to precision oncology. The role of oncologists is central to success as is working with multidisciplinary clinical teams that include geneticists, radiologists, pathologists, data scientists, and nurses. The key role of regulators adds an extra level of complexity as do ethical concerns.
AI solutions are often working on different image formats and different proprietary technology. 2 Since many AI systems are proprietary, compatibility and integration are also daunting issues. This is exacerbated by the lack of collaboration across hospitals and health maintenance organizations who sometimes view one another as competitors—instead of allies. The extent of collaboration needed across many disciplines—including with regulators—is unprecedented. While AI is moving forward rapidly, the development of governmental regulations to date has been slow.
Finding and retaining workers with data skills represents yet another challenge. The AI enabled environment requires not just a change in leadership mindset, but also more collaborative young and talented researchers. There is a significant shortage of talent and as AI adoption grows, so will the demand for talent. Precision oncology can learn from other industries where some companies have had success by hiring new graduates with bachelor’s and master’s degrees in computer science or engineering and training them on local issues. 3
Arguably the high cost of cost of integrating AI into precision oncology is one of the most daunting challenges. Developing sophisticated AI algorithms tailored for precision oncology involves significant effort and cost. Then, as AI in precision oncology relies heavily on data from sources like medical imaging and genomic sequencing—the gathering and cleansing of such datasets are often expensive. AI-driven tools for precision oncology must often go through regulatory approval (e.g., FDA or EMA approval) to ensure safety. These can also be lengthy and expensive. Once developed, many AI tools are based on proprietary software which may require additional expense in sharing across hospitals or cancer treatment centers.
While overcoming challenges related to risk averseness, data quality, collaboration, regulation, cost, and skills development are indeed formidable, the potential benefits of deploying AI in precision oncology are more than sufficiently compelling to justify the needed effort. These benefits include; but are not limited to, enhanced diagnostic accuracy, personalized treatment plans, improved treatment outcomes, predicting treatment outcomes, reduced trial-and-error in therapies, and more efficient use of resources—all of which contribute to more effective, personalized cancer care.