Oral cancer is the most frequent cancer in Indian men. It is also the second leading cancer in Indian women, next to breast cancer. Worldwide, the incidence of lip and oral cavity cancer cases is projected to rise from nearly 3.9 lakh new cases in 2022 to around 6 lakh by 2045. During this period, deaths due to oral cancers are expected to rise from 1.9 lakh to 2.94 lakh annually. Oral cancer now ranks as the 13th most common cancer globally.
Tobacco consumption has been the major causal factor driving the high incidence. Tobacco smoke contains over 70 known cancer-causing chemicals. They include polycyclic aromatic hydrocarbons and heavy metals like arsenic and cadmium. When these chemicals enter the cells of the mouth’s mucosal lining, they damage their genetic material and trigger abnormal cell growth. Smokeless oral tobacco products, too, release these carcinogenic chemicals. Chewing of tobacco products also continues to be a public health concern in India, despite several states banning the sale of ‘zarda’.
Areca nut (betel nut) is also a known cause of oral cancer. When consumed raw, dried or mixed in commercial preparations, areca nut can cause sub-mucous fibrosis (stiffening of the mouth tissues). This frequently turns into cancer, even when the nut is consumed without tobacco. Among other causes of oral cancer are: alcohol consumption; poor oral hygiene which permits growth of inflammation causing bacteria (dysbiosis); chronic mechanical irritation of the oral mucosa by a jagged tooth.
India has included screening for oral cancer in the primary healthcare component of Ayushman Bharat, the flagship national health programme. This is alongside screening for breast and cervical cancers. While the oral cavity is easy to inspect visually, detection of pre-malignant and malignant lesions in that space is often missed in primary care settings. Differentiation between non-cancerous, pre-cancerous and cancerous lesions is also not easy at that level of care. As a result, referral to higher levels of care for confirmatory diagnosis and initiation of appropriate treatment is delayed inordinately. This adversely affects survival rates.
Artificial intelligence has recently stepped into this arena, to improve the rates of early and accurate diagnosis. In recent years AI has found increasing application in the detection of organ and tissue pathologies in many disease conditions—even in a pre-clinical stage in several instances. With AI techniques rapidly advancing, clinical applications have grown manifold. Oral cancer too has seen many AI tools being used in research studies, to detect ‘potentially malignant’ as well as definitely cancerous pathologies of the oral cavity.
Convolutional neural networks have been most often used for this purpose. Deep convolutional neural networks, artificial neural networks, Random Forest and Decision Trees are among the other AI tools employed for oral cancer detection. Data inputs were obtained from photographs, mobile images, radiographic images and cytology of saliva samples. Recent studies have also included breath analysis for volatile organic compounds, generated by cancerous lesions.
AI-integrated imaging modalities have combined fluorescence lifetime imaging with end-to-end neural networks for in-vivo diagnosis. Cloud-based multi-spectral imaging was employed for real time detection.
Many of the studies evaluating the diagnostic accuracy of AI techniques for detection of oral cancer have reported levels of sensitivity and specificity around 90 percent. These indicate low rates of ‘false negative’ and ‘false positive’ results respectively. While these are encouraging, it must be recognised that these sensitivity and specificity rates only report how the tests performed overall in a mixed group of already diagnosed patients and non-patients. They do not help healthcare providers in assessing whether an individual sitting in front of them has disease or not.
Those probabilities are estimated by another set of derived numbers, called the ‘positive’ and ‘negative’ predictive values. These values are more pertinent to individual diagnosis. However, they are very vulnerable to the prevalence of the disease in the population. Lower the population prevalence, greater the likelihood of a ‘false positive’ value emerging. Context matters!
Better measures of a test’s diagnostic accuracy, which are not vulnerable to variations in population prevalence, are the ‘likelihood ratios’. Likelihood ratio of a positive test result can be derived by calculating sensitivity/1-specificity, while the likelihood ratio of a negative test result is derived from 1-sensitivity/specificity.
These ratios have relevance to the application of the Bayes’ theorem of ‘conditional probability’. It states that the post-test probability of disease is a product of the pre-test probability and the likelihood ratio derived from the test result.
AI diagnostics are mostly based on applications of Bayes’ theorem. Diagnostic confidence emerges from a hierarchy of increasing probabilities where a post-test probability estimated from an initial test becomes the ‘prior’ for the next level of testing. Research into the use of AI for detecting oral cancer must apply these statistical methods to advance the diagnostic value from the broader population to the individual person.
Even as we evaluate diagnostic accuracy through better statistical tools, there is good reason to believe that AI can be a valuable ally in the diagnosis and characterisation of pre-cancerous and cancerous lesions of the oral cavity. It can be applied cost-effectively in low resource settings and is potentially scalable. However, the ultimate proof of benefit from AI-guided diagnostic techniques will come from timely and effective clinical interventions catalysed by the test results.
What needs to be ensured is an adequate capacity of the health system to efficiently integrate such AI-driven diagnostic technologies at various levels of healthcare. This requires training of healthcare providers in applying and interpreting AI-based diagnostics. Once the initial diagnosis is made in primary care, there must be efficient triage for speedy referral to centres of advanced care for diagnostic confirmation and initiation of appropriate treatment.
The systemic capacity for providing such care must be enhanced at district hospitals and medical colleges. Only then will AI serve the purpose of improving the survival prospects of oral cancer patients.
K Srinath Reddy | Chancellor, PHFI University of Public Health Sciences; and Chair, Centre for Universal Health Assurance, Indian School of Public Policy
(Views are personal)
(ksrinath.reddy@phfi.org)