Transforming Healthcare through a new era of Artificial Intelligence and Personalized Medicine

-by Ketan Paranjape, Ph.D., MBA

Introduction

21st-century healthcare professionals are confronted by many technological advancements and large amounts of data. Physicians and nurses are overwhelmed by data from infusion pumps, vital sign monitors, laboratory tests, molecular tests, medical images and all the data that has been recorded in electronic medical records. Gathering this data and using it to make an informed and personalized decision poses a unique challenge that has yet to be overcome. New technologies such as artificial intelligence (AI) have the intrinsic ability to gain insights from large amounts of data from various sources and may be used to solve these problems.

Digitization of healthcare

An explosion of data and knowledge in medicine, diseases and science is beginning to impact the healthcare industry, bringing with it a real transformation in care. In the United States, the Health Information Technology for Economic and Clinical Health Act (HITECH Act) of 2009 resulted in Electronic Health Record (EHR) adoption to increase from 9.4% in 2008 to 83.8% in 2015 through financial incentives and increased penalties for violation of the HIPAA privacy and security rules [1].

Along with digitized medical records, it is estimated that by 2020, medical knowledge will double every 73 days [2]. A doctor would need to spend 29 hours a day absorbing new medical knowledge to stay up to date. In other words, we have reached the capacity of the human brain and time to follow and process the new medical knowledge that is being generated and published.

In an era of digital technology, we will be able to increasingly tailor medical treatment to the needs of individuals and small groups of patients. More information will be captured, stored, and analyzed to learn how diseases manifest themselves and how patients experience them every day. Combined with a deeper understanding of molecular science and new methods for diagnostics, this development will bring disruptive change to how we research, develop, approve, and pay for medicines, as well as how patients and their physicians make decisions about whether, when and how to treat their illnesses.

New sources of data

As knowledge in medicine, diseases and science grows, high-quality data from a wide array of sources can be collected for each patient and can be connected to data from large pools of other patients for analysis. This enables us to arrive at a deeper understanding of disease biology and its expression in individual patients. Patients are more knowledgeable and informed, and in a position to demand innovative and effective treatments. Real-world evidence, molecular information generated from next-generation sequencing, data from wearable devices and mobile apps and novel clinical trials are increasing our understanding of health and disease. The regulatory environment needs to and is evolving and adjusting for these novel approaches to healthcare. The task of unlocking the ecosystem of digital healthcare cannot be done by anyone alone. As a result, new types of partnerships are forming to ensure we are moving towards value-based, personalized patient care.

AI in healthcare

With the digitization of healthcare, technologies such as AI can help us analyze these vast amounts of data to derive insights and help with decision-making.

AI in healthcare is the use of complex algorithms and software to emulate human cognition in the analysis of complicated medical data without direct human input. Since a seminal paper by Sir Alan Turing in 1950 [3], AI has had many advances in Natural Language Processing (NLP), Machine Learning [4], Deep Learning [5], Speech Recognition, Virtual Agents, and AI-optimized Hardware, amongst others.

Today, AI is already used in healthcare [6] for example to decrease false-positive results in screening for breast cancer [7], reduce medical transcription costs [8], and improve physician workflow while relieving and helping to prevent burnout [9], robotic surgery resulting in a shorter length of hospitalization and loss of blood and predicting mortality rates of patients with acute heart failure.

In the past, the most important stakeholder in healthcare, which is the patient, suffered from a broad category of diseases which were treated with the same medicines, leaving physicians to puzzle over why they worked for some people and not others.

Today scientists have begun to understand, target, and diagnose illnesses on an individual level and AI can play a significant role in this process given its unique capabilities of detecting subtle disease-specific patterns from a wide array of sources, such as molecular diagnostics, that humans would never recognize.

Personalized medicine

With the use of machine learning applications, a subcategory of AI, that can combine data from all state-of-the-art diagnostic tests and other resources, there is more potential for personalized medicine than ever before. A high-level discussion of two specific fields of medicine will show what AI, in combination with all these recent technologies, can and cannot do.

Lung Cancer

A 2018 narrative review on AI applications for non-small cell lung cancer shows that there are already many applications being tested in this field [10]. Machine learning algorithms can be used to increase our understanding of important genomic pathways in lung cancer, with the use of microarray data. Also, machine learning can be used to predict which patient will respond to newly developed checkpoint inhibitors or personalised radiation therapy, thereby choosing an optimal treatment strategy. A key feature in the success of AI for lung cancer is that many molecular abnormalities have already been discovered, such as mutations in the epidermal growth factor receptor (EGFR) and anaplastic lymphoma kinase (ALK). These extremely specific markers provide an excellent starting point for algorithms to work from.

Sepsis

A similar narrative review of AI applications for sepsis was published in 2019 [11], showing that applications to improve diagnosis, treatment and prognosis exist already. Many algorithms to predict sepsis onset have been developed, with encouraging results. However, there are no clear molecular abnormalities on which new algorithms can be trained. The rapid onset and heterogeneous presentation of this syndrome make it so, that the understanding of pathophysiology remains poor when compared to that of lung cancer. The potential of AI is therefore limited, as unique features needed to do adequate predictions are not yet known. Machine learning can classify in the absence of unique features, but to detect conditions such as sepsis more data is needed because of heterogeneous presentation and unique features that are needed to provide understanding to develop new treatments.

Algorithms can be trained to predict the best possible treatment on an individual level but can only consider the general treatments that exist today - antibiotics, source control and intravenous fluids. Likely, better treatment options exist, but machine learning algorithms are limited by human knowledge at this point in time.

For AI to be able to provide personalized predictions for treatment, meaningful data at scale is needed. Clinical trial data, molecular data and general patient data need to be integrated into advanced predictive models. A broad understanding of pathophysiology in a certain field is needed in order for AI to become valuable.

Mainstreaming AI in healthcare

As discussed, some disease-specific challenges, such as with sepsis, hold back the mainstream adoption of AI in certain fields for now, but there are also some general concerns and challenges about the adoption of AI in healthcare which must be addressed at a larger scale.

Challenges with AI

Challenges with the introduction of AI in healthcare are centred around explainability [12], liability [13] and privacy. Furthermore, the medical education system for healthcare professionals will have to undergo a rigorous transformation.

The lack of explainability of AI algorithms is likely to bring about some resistance from the medical community. The more accurate the algorithms, such as neural networks, the less explainable they are. This “black box” phenomenon makes it hard for healthcare professionals to get used to working with AI and trusting the algorithm [14]. In the end, physicians still must make a final decision and not knowing why you would make a certain decision will raise many more issues when a patient is given the wrong diagnosis. Software developers will have to take this into account and prioritize both explainability and accuracy. Having explainability will also simplify acceptance by the US Food and Drug Administration (FDA) as mentioned in the recent documents regulating AI [15] and by medical regulatory agencies around the world. 

Then there is the issue of liability. Who is to blame when something goes wrong? There is no case law about the use of medical AI yet. Even worse, the current laws seem to incentivize physicians to minimize the potential value of AI as they will only face liability when current protocols are not adhered to. New malpractice laws will have to be developed to specify the liability of all involved parties: healthcare professionals, hospitals, software companies, software developers and data collectors.

Privacy is another outstanding issue with the use of AI. Vast amounts of patient data are needed for some AI algorithms to properly function. Google for example is using 46 billion data points collected from 216,221 adults’ de-identified data over 11 combined years from two hospitals to predict the outcomes of hospitalized patients [16]. This raises questions about how this data is obtained and whether all patients have had a fair chance to decide about the use of their data.

Lastly, as patients begin to see the benefits of AI and proactively use tools such as chatbots, physicians will need to be aware of the limitations of such technologies and care for the patient accordingly. They will need to be trained in how to effectively use such technologies to their benefit and help ease their burden.

Resolving Challenges

To alleviate the main concern with explainability we need models that can explain the why, so a physician can confidently diagnose a patient with a certain disease. Explainable AI is a new emerging discipline that is working towards making machine decisions transparent, interpretable, traceable, and reproducible [17].

The healthcare community needs to be educated regarding these challenges and how to address them and establish standards and guidelines so a physician and a machine working together have the greatest potential to improve clinical decision-making and patient health outcomes.

Medical students, residents, fellows and practising physicians need to have knowledge of AI, data sciences, EHR fundamentals and ethics and legal issues concerning AI. Medical schools will need to include them as part of the curriculum. A staged approach to educating medical students through their journey is recommended [18].

In Jun 2018, the American Medical Association’s House of Delegates comprised of proportional representations of every major national medical speciality society and state medical association adopted its first policy on healthcare Augmented Intelligence[19]. Some of the recommendations included identifying opportunities to integrate the perspective of practising physicians into the development, design, validation and implementation of healthcare AI; encouraging education for patients, physicians, medical students, other healthcare professionals, and health administrators to promote a greater understanding of the promise and limitations of healthcare AI; and exploring the legal implications of healthcare AI, such as issues of liability or intellectual property, and advocate for appropriate professional and governmental oversight for safe, effective, and equitable use of and access to healthcare AI.

Conclusions

Both personalized medicine and AI are evolving. As we understand more about biology, diagnostics, and augment medical knowledge with patient data from images, genomics, and medical records, we will be able to identify personalized therapies for individuals. As we gain a deeper understanding of how AI works, healthcare professionals will be able to explain the decision they make with the help of AI tools. With the help of technology and regulatory bodies, we will be able to resolve challenges with liability and privacy. We are well on our way to providing personalized treatment strategies driven by AI.

References:

[1] “Adoption of Electronic Health Record Systems among U.S. Non-Federal Acute Care Hospitals: 2008-2015.” [Online]. Available: https://dashboard.healthit.gov/evaluations/data-briefs/non-federal-acute-care-hospital-ehr-adoption-2008-2015.php. [Accessed: 25-Nov-2019].

[2] P. Densen, “Challenges and Opportunities Facing Medical Education,” Trans. Am. Clin. Climatol. Assoc., vol. 122, p. 48, 2011.

[3] “Computing Machinery and Intelligence on JSTOR.” [Online]. Available: https://www-jstor-org.iclibezp1.cc.ic.ac.uk/stable/2251299?seq=1#metadata_info_tab_contents. [Accessed: 28-Jul-2019].

[4] B. J. Erickson, P. Korfiatis, Z. Akkus, and T. L. Kline, “Machine Learning for Medical Imaging,” RadioGraphics, vol. 37, no. 2, pp. 505–515, Mar. 2017.

[5] R. Miotto, F. Wang, S. Wang, X. Jiang, and J. T. Dudley, “Deep learning for healthcare: review, opportunities and challenges,” Brief. Bioinform., vol. 19, no. 6, pp. 1236–1246, Nov. 2018.

[6] K.-H. Yu, A. L. Beam, and I. S. Kohane, “Artificial intelligence in healthcare,” Nat. Biomed. Eng., vol. 2, no. 10, pp. 719–731, Oct. 2018.

[7] J.-J. Mordang, A. Gubern-Mérida, A. Bria, F. Tortorella, G. den Heeten, and N. Karssemeijer, “Improving computer-aided detection assistance in breast cancer screening by removal of obviously false-positive findings,” Med. Phys., vol. 44, no. 4, pp. 1390–1401, Apr. 2017.

[8] K. Saxena, R. Diamond, R. F. Conant, T. H. Mitchell, I. G. Gallopyn, and K. E. Yakimow, “Provider Adoption of Speech Recognition and its Impact on Satisfaction, Documentation Quality, Efficiency, and Cost in an Inpatient EHR.,” AMIA Jt. Summits Transl. Sci. proceedings. AMIA Jt. Summits Transl. Sci., vol. 2017, pp. 186–195, 2018.

[9] “3 ways medical AI can improve workflow for physicians | American Medical Association.” [Online]. Available: https://www.ama-assn.org/practice-management/digital/3-ways-medical-ai-can-improve-workflow-physicians. [Accessed: 03-Feb-2019].

[10] M. Rabbani, J. Kanevsky, K. Kafi, F. Chandelier, and F. J. Giles, “Role of artificial intelligence in the care of patients with non-small cell lung cancer,” European Journal of Clinical Investigation, vol. 48, no. 4. Blackwell Publishing Ltd, 01-Apr-2018.

[11] M. Schinkel, K. Paranjape, R. S. N. Panday, N. Skyttberg, and P. W. B. Nanayakkara, “Clinical applications of artificial intelligence in sepsis: A narrative review,” Comput. Biol. Med., p. 103488, Oct. 2019.

[12] A. Holzinger, C. Biemann, C. S. Pattichis, and D. B. Kell, “What do we need to build explainable AI systems for the medical domain?,” Dec. 2017.

[13] “Artificial Intelligence, Medical Malpractice, and the End of Defensive Medicine | Bill of Health.” [Online]. Available: https://blogs.harvard.edu/billofhealth/2017/01/26/artificial-intelligence-medical-malpractice-and-the-end-of-defensive-medicine/. [Accessed: 14-Oct-2018].

[14] R. Miotto, L. Li, B. A. Kidd, and J. T. Dudley, “Deep Patient: An Unsupervised Representation to Predict the Future of Patients from the Electronic Health Records,” Sci. Rep., vol. 6, no. 1, p. 26094, May 2016.

[15] “Artificial Intelligence and Machine Learning in Software as a Medical Device | FDA.” [Online]. Available: https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-software-medical-device. [Accessed: 09-Dec-2019].

[16] “Google is using 46 billion data points to predict a hospital patient’s future — Quartz.” [Online]. Available: https://qz.com/1189730/google-is-using-46-billion-data-points-to-predict-the-medical-outcomes-of-hospital-patients/. [Accessed: 01-Nov-2018].

[17] L. H. Gilpin, D. Bau, B. Z. Yuan, A. Bajwa, M. Specter, and L. Kagal, “Explaining Explanations: An Overview of Interpretability of Machine Learning,” May 2018.

[18] K. Paranjape, M. Schinkel, R. Nannan Panday, J. Car, and P. Nanayakkara, “Introducing Artificial Intelligence Training in Medical Education (Preprint),” JMIR Med. Educ., Sep. 2019.

[19] “AMA Passes First Policy Recommendations on Augmented Intelligence | American Medical Association.” [Online]. Available: https://www.ama-assn.org/ama-passes-first-policy-recommendations-augmented-intelligence. [Accessed: 15-Oct-2018].