To address the twin problems of diagnosing cancer early and to treat it at a fraction of the cost, the field of Artificial Intelligence must be explored
In 1910, the then US President, William Howard Taft, had claimed that “within five years, cancer will have been removed from the list of fatal maladies.” More than a century later, we are yet to understand the intricacies of the disease. Cancer refers to a class of disease where previously healthy cells multiply abnormally and spread or “metastasise” to other parts of the body. When this growth impairs the normal functions of organs or systems, it can lead to death. Cancer is not a single disease but a grouping of hundreds of diseases, which share common features. This means that any cure must be specific to the particular sub-class to be effective. This, along with the fact that curing cancer ultimately means removal of unwanted growths of parts of our own body, make it incredibly hard to treat.
Cancer is of particular concern for India. There has been a tectonic shift in the disease burden faced by the country. Vector-borne diseases, such as that of Malaria, have steadily decreased over the last three decades but the incidence of lifestyle diseases, categorised as non-communicable diseases or NCDs, such as diabetes and hypertension, has increased manifold in the same period. Increasingly, cancer’s share in NCDs has increased in the past few years. The age-standardised rate of cancer is estimated to be 97 per 100,000 people. Treatments are prohibitively expensive and out-of-pocket expenditure is highest among any ailment, with about 40 per cent of cancer cases being financed through borrowings, sale of assets and contributions from family.
In 2014, the ratio of the number of oncologists to the number of cancer patients stood at 1:2000, a far cry from the 1:100 ratio of high income nations such as the US. With 11.5 lakh new cancer patients being registered every year in the country, the healthcare system is reeling under overburdened workforce and inadequate infrastructure. In 2018 alone, 7.8 lakh people succumbed to the disease. Cancer takes time to present itself, which means it is usually diagnosed at an advanced stage. This inevitably means that the disease is tougher and more expensive to treat. Early diagnosis, coupled with prompt treatment, especially in the case of common cancers such as breast, cervical and colorectal cancers, perhaps is the only way forward in ensuring survivability. But diagnostic services for cancer are scant, the procedures tend to be expensive and invasive. The problem India faces is two-fold: We need to diagnose cancer earlier and do it at a fraction of the cost. Artificial Intelligence (AI) may just be the answer.
Field image recognition, a sub-part of AI, has the ability to read and point anomalies. Many companies have come forward and developed AI-enabled radiology tools that predict cancer at an early stage. Existence of a high number of digitised images in radiology makes it a ripe field for AI exploration. Emergence of algorithms with the capability of analysing digital images to individual pixel level has opened the possibility of detecting features imperceptible to the human eyes.
MIT’s AI lab and Massachusetts General Hospital (MGH) have created a new deep-learning model that can predict from a mammogram if a patient is likely to develop breast cancer as much as five years in the future. Identifying subtle markers in breast tissues that can act as precursors to malignant tumors, years in advance, is a breakthrough that supports physician’s clinical decision. Niramai Health Analytix, a startup headquartered in Bengaluru, has developed an AI-led diagnostic platform that uses thermal image processing and ML algorithms for reliable and accurate breast cancer screening.
The tech community has recognised the impact that AI can have in healthcare with open source community and medical establishments actively building data-sets to spur innovation. Detection of lung nodules and their classification into benign and malignant growths using CT scans formed the problem statement of 2017 Kaggle Data Science Bowl, an international competition in the field of machine learning. Responses to this resulted in models with promising accuracy ranging from 85-90 per cent. AI can also successfully segment the lung tumors based on volumes and offer an insight into triaging cases for doctors and radiotherapy treatment planning. Whether these algorithms can be deployed on field, after refinement, is a question the medical community and the Government needs to answer.
AI may, perhaps, already be better than human doctors in a few, very specific tasks. Taking the conversation outside oncology, NTT Data, a Japanese technology firm, partnered with Deenanath Mangeshkar Hospital and Research Center in Pune to test the efficacy of its AI diagnosis support solution. The model was able to detect 56 emphysema cases, a lung condition that causes difficulty breathing, while normal diagnosis without AI detected 17 cases. These AI detected cases showed signs of mild or moderate levels of emphysema, which gave early findings to kickstart early treatment, thereafter enabling tracking of disease progression. In oncology as well, AI may be able to diagnose patients earlier and more accurately than doctors can, with the tools they currently have. Applications of AI in oncology go beyond image recognition as well.
With the healthcare industry increasingly adopting the practice of maintaining digitized health records, referred to as Electronic Health Records (EHR), AI is increasingly harnessing these digital records using natural language processing techniques to analyse data and predict the development of diseases. Marriage of EHR and AI find applications not only in disease prediction but also in disease monitoring, decision-making and drug recommendation. The use of EHR-enabled AI, to provide better services to the patients, is already underway in India.
As it stands, there is ample evidence to prove that AI will make diagnosis and treatment of cancer cheaper, more accurate and accessible to all. A few start-ups and healthcare institutions are well underway at making this promise a reality. With advances in technological infrastructure and burgeoning research in AI, we can begin to expect path breaking changes in oncological care. But lack of clear regulations, quality data and a disparate public health system are impediments to large-scale deployment of these solutions.
Perhaps the biggest barrier to the injection of these solutions in public health systems lies in proving generalisability and on ground efficacy of these applications. Some AI algorithms make it impossible to re-trace the steps involved in the algorithm reaching a decision, a problem which is referred to as “black box problem.” This problem might prevent us from weeding out biases and can result in unwitting and malicious problems. Biases may creep in through other mechanisms, such as training the AI model on non-representative datasets. Regulations, which mandate and set parameters for datasets that AI algorithms are trained on, are required to prevent any such biases. An ecosystem that supports infusion of AI into regular practices needs to be developed as well. As it stands, public healthcare systems are fragmented and are not interoperable. While a few standards do exist, the healthcare industry is yet to adopt them in a holistic manner. Adoption of industry wide standards along with clear mechanisms and guidelines for use and certification of AI and AI-enabled medical devices will be necessary in order to make sure AI solutions for cancer become ubiquitous.
(Eluri has led digital transformation projects in the pharmaceutical industry. Sharma is a software professional. Both are with the International Innovation Corps, Chicago University)
