AI in Healthcare White Paper

 

 

 

This article was written by Amber Zhou and Kwon Sok Oh, Financial Analysts at I Know First.

 

Highlights

Abstract

Part One: Artificial Intelligence in Healthcare

  • Overview of AI in Healthcare
  • Debate on Whether AI would replace doctors
    • AI Could Replace Major Doctor Roles
    • Argument on AI Replacing Only Some Major Doctor Roles
    • Argument on AI Replacing All Major Doctor Roles
  • Obstacles of AI in Healthcare

Part Two Epidemics Forecasting

  • Epidemics Forecasting
  • Surveillance System from Internet-based Sources
    • Different Methods
    • Accuracy and Timeliness
  • Comparison with Weather Forecasting
    • Similarities
    • Differences

Conclusion

I Know First’s Adaptation of AI in Healthcare

Abstract

AI has been exponentially changing modern human life during the past decade, and it is expanding its area of influence to various parts of people’s day to day lives. From a broad point of view, AI is being used in healthcare, especially within the realm of diagnosis, and epidemics forecasting. More specifically, we will discuss how AI is used for differential diagnosis, the extent to which AI would replace doctors, and the obstacles for AI to be used in healthcare. Finally, we will discuss various methods of surveillance systems using AI to forecast epidemics and compare the effectiveness to weather forecasting.

Part One: Artificial Intelligence in Healthcare

1. Overview of AI in Healthcare

AI in healthcare is the use of artificial intelligence to approximate human decision making in the healthcare industry, including diagnostics, treatment process outlining, drug development, personalized medicine, and patient monitoring. Various algorithms and software are used to analyze complex medical data and make accurate conclusions without explicit human input. The primary area of AI used in healthcare is the analysis between disease prevention or treatment and the respective outcomes. Major medical institutions, including the Mayo clinic and Massachusetts General Hospital, as well as major tech companies, including Google and IBM, are leading the development of AI in healthcare by developing various algorithms and software that can be used in the healthcare industry.

2. Debate on Whether AI Would Replace Doctors

There is widespread debate as to whether AI in healthcare would replace the roles of doctors. Experts in the field are divided among arguments that AI would not be able to replace doctors, AI would replace some roles of doctors, and AI would completely replace doctors. In the frontier of this debate is radiology due to its unique data intense characteristics. Most experts in AI and radiology speculate that AI would be able to analyze all relevant clinical images along with extensive patient medical history. Furthermore, AI will not be affected by fatigue, monotony, and distractions. The debate is whether such supreme performance metrics would allow AI to all major doctor roles or just some major doctor roles. Moreover, there is a middle ground argument that AI would replace all major doctor roles, but not in the foreseeable future relevant to students currently in medical school.

a. AI Could Replace Major Doctor Roles

Proponents of the argument that AI will replace major doctor roles include Dr. Michael Recht of NYU Langone Health and R. Nick Bryan of the University of Texas in Austin. They speculate that in the near future, roughly 10 years from now, AI would play of dominant part in radiology, aiding doctors to be more accurate and efficient. AI is predicted to perform extensive pre-analysis of medical images before radiologists actually evaluate them. This will allow radiologists to spend their time more efficiently by separating items on medical images that need urgent attention and those that do not. AI would also be able to perform mundane, yet mandatory routine tasks, such as quantification, segmentation, and pure pattern recognition. If AI does indeed replace such major tasks of radiologists, then radiologists will have more time to invest in “human roles,” such as interacting more with patients and playing more integral roles in integrated clinical teams. Dr. Seth Berkowitz of Beth Israel Deaconess Medical Center points out that AI lets radiologists to prove that what they do is encompass more that just translating medical images into radiology reports.

b. Argument on AI Replacing Only Some Major Doctor Roles

So the questions is, how much of these major roles of radiologists will be replaced? Dr. Berkowitz thinks that although some major radiologist roles would be replaced by AI, it would be impossible for all major roles to be replaced. In the paper he published on JACR, he said that radiologists will not be replaced by machines and that radiologists of the future would also take up roles as data scientists of medicine.

c. Argument on AI Replacing All Major Doctor Roles

On the other hand, some AI and radiology experts think that sooner or later, AI would completely replace all major roles of radiologists. Dr. Jason Kelly of Sky Ridge Medical Center argues that the sheer processing power of AI will allow it to outpace human radiologists, allowing it to analyze more studies and medical images without human intervention. The skyrocketing demand for medical images and studies to be analyzed is driving the tides in favor of AI against radiologists. Moreover, the increase of the number of scans and slices will for emergency room patients is another favorable factor for AI developers, not to mention the cost efficiencies of machines compared to human beings. Dr. Kelly points out that just like cars replaced all major roles of horses, AI will replace all major roles of radiologists.

3. Obstacles of AI in Healthcare

Regardless of the conclusion of such debate, AI in healthcare has a number of obstacles to get through before becoming fulling integrated into the world of medicine. AI analysis and diagnosis will need to get past bureaucracy approval, including that of the FDA, which is known for its rigor. Furthermore, patients are going to need persuasion to allow AI to take up the roles of traditional human doctors. Older patients unfamiliar with technology will have a hard time accepting that AI could do what their previous doctors have done without much downside. Finally, AI software developers will have to constantly be aware of potential lawsuits regarding the mistakes that their products make.

Most importantly, for each radiologist task, customized AI will have to be developed accordingly. Currently, there are AI’s that prioritize worklists, point out anatomical abnormalities, and calculate probabilities regarding critical findings, but this is only a small portion of the numerous kinds of AI that will have to be developed in the future.

Moreover, there is somewhat of a consensus that no matter how developed AI gets, there must be some kind of final human checking procedures. Whatever the conclusions that AI in radiology makes, human radiologists should do a final review in order to assure that the AI is operating properly, and this step seems to be inevitable for the moment. Whether this final human review step will be necessary in the future is up to how intricate the developed AI will be.

Finally, the general culture of medicine incorporates an unwillingness to accept and embrace human uncertainty. Even though physicians are rationally aware that human decision making is susceptible to uncertainly, the integration of symptoms and test results in order to create a diagnosis is deeply rooted into the medical society. Such process causes oversimplification, and information lost in this way inevitably creates uncertainty in the final diagnosis. On the other hand, AI does tolerate uncertainty, so instead of giving a black and white diagnosis like human doctors, AI is able to give a gray-scale diagnosis encompassing all possible outcomes.

Part Two: Epidemics Forecasting

1. Epidemics Forecasting

An epidemic is defined as the rapid spread of infectious disease to a large number of people in a given population within a short period of time, usually two weeks or less (U.S. Department of Health and Human Services, 2012). Historically, research efforts were focused on analyzing the spread and control of epidemics. However, an early and rapid detection of an epidemic outbreak plays an important role in facilitating timely medical intervention, reasonable public health resource distribution and mitigating the disease’s impact. The epidemics forecasting filed is attracting more attention, with attempts to answer questions such as “When will the disease outbreak to a peak?” and “How much of the population will suffer from this disease in a month?” (Moran et al., 2016). Various mathematical models have been developed to generate both short-term and long-term epidemics prediction. They are built to capture the characteristics of early epidemics growth patterns from both historical and simulation infectious disease outbreak data (Chowell, Sattenspiel, Bansal & Viboud, 2016). Besides the traditional mathematical methods, machine learning and artificial intelligence are broadly involved for more accurate forecast in a variety of methods illustrated below.

2. Surveillance System from Internet-Based Sources

a. Introduction

According to the definition by the World Health Organization (WHO), public health surveillance is “The continuous, systematic collection, analysis and interpretation of health-related data needed for the planning, implementation, and evaluation of public health practice” (“Public health surveillance”, 2018). The main purpose of developing surveillance system is to minimize the morbidity and mortality rate caused by epidemic diseases through early detection and real-time trend monitoring regarding the disease size, spread and tempo of outbreaks.

Traditional surveillance methods focused on monitoring validated public health indicators, such as mortality, morbidity, clinical data, surveys, animal studies and demographic data. However, there exists a fatal drawback for these methods in timeliness. There is usually significant time lag between the beginning of an epidemic outbreak and official public data normally due to tedious validation processes and bureaucratic barriers. As a result, the detecting and responding efficiency would be greatly damaged (Yang, Horneffer & DiLisio, 2013).

To mitigate the impacts brought by delay in data collection process, an alternative Internet-based method was developed. As we are entering the era of big data, a huge amount of unofficial data from all kinds of Internet sources, including online news, blogs and social media, is easily and quickly accessible. Hence this new method may largely address the timeliness problem by simply identifying events or infectious disease-related keywords online (Chunara, Freifeld & Brownstein, 2012).

b. Three Streams

According to a study review conducted by Yan, Chughtai & Macintyre, Internet-based surveillance methods could be categorized into three groups by the data sources used: (i) existing surveillance systems and news aggregators, (ii) search query surveillance and (iii) social media surveillance.

Exhibit: Internet-based syndromic surveillance methods

(Yan, Chughtai & Macintyre, 2017)

Existing surveillance and news aggregators are event-focused systems collecting unstructured data from news reports and official released data (Yan, Chughtai & Macintyre, 2017). The collection process is mostly done by automated tools, except that some manual corrections and judgements are involved occasionally. Examples include HealthMap, BioCaster and ProMED-mail.

Search query surveillance is built on the epidemic-related information requesting behavior demonstrated on search engines. Data are either publicly available, such as Google Flu Trends, or obtained from the search engine provider (Woo et al., 2016). Under this method, the keyword searching frequencies or volumes would be studied and validated in conjunction with pre-existing official data and relevant statistical models (Yan, Chughtai & Macintyre, 2017).

Social media surveillance focuses on analyzing posting content on different social media platforms, including the online restaurant websites and blogs. In addition, the model could be further extended by extra information about the user names, message types, users’ number of followers, IP addresses and their correlations (Odlum & Yoon, 2015). Dataset templated could be customized based on which country and which social media platform we are studying.

c. Timeliness and Accuracy

If we use official data and public health reports as benchmarks, the timeliness of all three types of surveillance systems get improved by varying degree. According to studies conducted up to now and summarized by Yan, Chughtai & Macintyre, existing surveillance and news aggregators provided epidemic alerts earlier than WHO reports by a minimum of 1.9 days and a maximum of three months. For search query surveillance, the time advance ranges from one to six weeks. Surveillance systems using intelligence from social media data issue earlier alerts than official reports by one hour to one week.

In terms of accuracy, Yan, Chughtai & Macintyre reviewed the correlations between different surveillance systems and official data evaluated in researches and studies. For all three types of systems, the results are generally positively correlated with official data but with large variations depending on the specific surveillance and variables chosen.

3. Comparison with Weather Forecasting

a. Similarities

The construction of epidemic forecasting models has drawn a lot from the experience of weather forecasting due to the essential underlying similarities. More specifically, both models are built on real-time continuous stream of data, one from weather observations and one from clinical or Internet-based data. Hence nonlinear dynamic models, data assimilation and uncertainty quantification in the process are needed (Moran et al., 2016).

b. Differences

Several major differences between the two cases make the application of weather forecast on epidemic forecast difficult (Moran et al., 2016). First, while the two models are both dynamic, weather models are easier to build because they are based on relatively unchanged physical principles. This is not the case for epidemic models since they tend to change with individual social behavior, which is much harder to predict. Secondly, data for weather forecasting models is much more objective, accurate and accessible compared with that for epidemic models, especially at the early stage of diseases. Thirdly, unlike weather forecasting, there is no established or consistent way of communicating the epidemic predicting uncertainty to the public or decision makers. This would impact users’ understanding and interpretation of predicting results.

4. Conclusion

AI is certainly being more and more integrated into the daily lives of human beings, but the extent to which AI can be fully utilized is still uncertain. AI is starting to be used in healthcare, especially for differential diagnosis. However, the extent to which AI will replace doctors is still widely debated, and the obstacles hindering AI to be applied extensively into healthcare remain unsolved.

Furthermore, epidemic forecast is of great use in facilitating timely infectious disease detection, trend monitoring and treatment intervention. As technology develops and more software tools are available, this field is under great attention and also experiencing rapid growth with the emergence of a variety of methods based on different data source. At the same time, it is still facing serious challenges, mainly from the changing human social behavior involved and the enormous but unstructured and messy data available.

5. I Know First’s Adaption of AI in Healthcare

I Know First has started to jump into the realm of AI in healthcare. Recently, I Know First’s CEO Yaron Golgher and KST Medical Group’s Medical Chairman Chong Han met up in person to discuss the partnership between the two companies for AI in healthcare. More information regarding this meeting can be found in previous articles through this link and this link.

References

Chowell, G., Sattenspiel, L., Bansal, S., & Viboud, C. (2016). Mathematical models to characterize early epidemic growth: A review. Physics Of Life Reviews18, 66-97. doi: 10.1016/j.plrev.2016.07.005

Chunara, R., Freifeld, C., & Brownstein, J. (2012). New technologies for reporting real-time emergent infections. Parasitology139(14), 1843-1851. doi: 10.1017/s0031182012000923

Kelly, J., (2016). The Robot Is In, And It Will See You Know. Forbes. url: https://www.forbes.com/sites/realspin/2016/02/11/automation-telemedicine/#2637b55b1775

Kruskal, J., Berkowitz, S., Geis, J., Kim, W., Nagy, P., Dreyer, K., (2017). Big Data and Machine Learning – Strategies for Driving This Bus: A Summary of the 2016 Intersociety Summer Conference. Journal of the American College of Radiology14(6), 811-817. doi: 10.1016/j.jacr.2017.02.019

Moran, K., Fairchild, G., Generous, N., Hickmann, K., Osthus, D., & Priedhorsky, R. et al. (2016). Epidemic Forecasting is Messier Than Weather Forecasting: The Role of Human Behavior and Internet Data Streams in Epidemic Forecast. Journal Of Infectious Diseases214(suppl 4), S404-S408. doi: 10.1093/infdis/jiw375

Odlum, M., & Yoon, S. (2015). What can we learn about the Ebola outbreak from tweets?. American Journal Of Infection Control43(6), 563-571. doi: 10.1016/j.ajic.2015.02.023

Public health surveillance. (2018). Retrieved from http://www.who.int/topics/public_health_surveillance/en/

Recht, M., Bryan, R., (2017). Artificial Intelligence: Threat or Boon to Radiologists? Journal of the American College of Radiology14(11), 1476-1480. doi: 10.1016/j.jacr.2017.07.007

Simpkin, L., Schwartzstein, M., (2016). Tolerating Uncertainty – The Next Medical Revolution, New England Journal of Medicine, 375, 1713-1715. Doi: 10.1056/NEJMp160642

U.S. Department of Health and Human Services. (2012). Principles of Epidemiology in Public Health Practice. Atlanta: Centers for Disease Control and Prevention (CDC).

Woo, H., Cho, Y., Shim, E., Lee, J., Lee, C., & Kim, S. (2016). Estimating Influenza Outbreaks Using Both Search Engine Query Data and Social Media Data in South Korea. Journal Of Medical Internet Research18(7), e177. doi: 10.2196/jmir.4955

Yan, S., Chughtai, A., & Macintyre, C. (2017). Utility and potential of rapid epidemic intelligence from internet-based sources. International Journal Of Infectious Diseases63, 77-87. doi: 10.1016/j.ijid.2017.07.020

Yang, Y., Horneffer, M., & DiLisio, N. (2013). Mining social media and web searches for disease detection. Journal Of Public Health Research2(1), 4. doi: 10.4081/jphr.2013.e4