Powering healthcare technologies with Artificial Intelligence
by Dhakshenya Dhinagaran
Artificial Intelligence (AI) refers to computer systems which are capable of performing tasks that would usually require human intelligence. Machine learning and deep learning functions allow such programs to sense, reason, act and adapt like humans. Machine learning, as the name implies, refers to algorithms whose performance improve as they are exposed to more data over time. Thus, like humans who mature with each new experience, these programs “learn” with every interaction they have with the user, without being specifically programmed every step of the way.
Like the brain, deep learning draws on the concept of neural networks which connect with each other in several layers. Information is extracted from each layer, to home in on a conclusion. For example, when identifying cell types, the first layer may analyse the cell membrane, the next layer, a cell specific organelle, and successive layers would continue to uncover cell specific characteristics until the cell is identified. The diagram (from cousins of AI) illustrates the relationship between AI, machine learning and deep learning.
While self-aware robots and machine networks like the Terminator and Skynet depicted in the Terminator movies are not reality yet, AI is already influencing daily life. For example, AI algorithms make recommendations on the movies that we watch and the food we eat by predicting our preferences. They also complete Internet searches and cleverly filter out most of the spam mail that we receive.
The ability of AI applications to seamlessly integrate into our lives and provide handy services suggests that they hold the potential to become more involved in our day-to-day activities. In fact, Google CEO Sundar Pichai described the involvement of AI in daily life as more profound than electricity or fire, accentuating the great potential of AI to be an incredibly disruptive form of technology.
Recently, AI has also been helping people with their shopping. Ever been frustrated when you don’t know the name of the item you want to buy? In such cases, don’t you wish you could just snap a picture of your desired product and pump it into the search engine? Well, thanks to AI, now you can! Amazon has implemented a visual search option on the smartphone application which enables you to view and buy items similar to the image you provide them with.
Fancy listening to a tune composed by AI? AI made its musical debut in 2017 when music producer Alex da kid consulted IBM Watson’s cognitive computing platform. The AI platform was first used to analyse big data comprising a list of billboard songs from the previous five years, social media commentaries, newspaper headlines, speeches and lectures. This helped assess the current mood and “emotional temperature” of the public, revealing “heartbreak” as an ideal theme for the new single. Watson’s machine learning algorithms were then employed to generate various musical excerpts for the producer to select from and work with. The resulting track, “Not Easy”, was well received and secured the number four position on the iTunes Hot Tracks chart in less than 48 hours of being released!
AI also has the potential to achieve feats beyond human capabilities. This was demonstrated in an online match of the board game “Go,” between AlphaGo, AI computer program trained to play ‘Go’ and Lee Sedol, an 18-time world champion professional ‘Go’ player. AlphaGo won by making an elegant move which was described by Lee as “not a human move”. This reiterates the ability for AI to unveil ideas, concepts and methods which humans on their own have yet to uncover.
The potential of AI to enhance human capabilities has been applied in the biomedical field. The Chinese startup Infervision proposed the use of deep learning to improve medical imaging analysis, with a specific goal to diagnose lung cancer – a condition with high mortality rates in China. From analysing medical images, the algorithm can pick out cancerous lung cells. This task which would normally take doctors 15-20 minutes, can be completed by Infervision in under 30 seconds!
The prediction of 3D protein structures from amino acid sequences is an especially challenging process which can benefit from the power of AI. To overcome the “protein folding problem”, DeepMind developed the AlphaFold software. AlphaFold emerged tops at the 13th Critical Assessment of Structure Prediction (CASP), a biannual competition aimed at predicting the 3D structures of proteins. An AI system which accurately predicts protein folding may be beneficial in understanding and tackling conditions caused by protein misfolding, such as Alzheimer’s, Huntington’s and Parkinson’s.
Applications in medicine and life sciences
Machine learning algorithms have also been especially beneficial in medical applications and life science research. They have been used in clinical trials to predict the efficiency of combination therapies, eliminating futile trial and error combinations.
With the increasing applications of AI in the medical field, the number of AI-based health initiatives funded globally rose from 20 in 2012 to around 70 in 2016. The union of AI with healthcare and biomedical innovations is encompassed within these four pillars: (1) therapeutics, (2) population health management, (3) administration and (4) diagnostics.
(1) Therapeutics – eg. Drug discovery
AI and machine learning have been cited as powerful tools to aid drug discovery, especially for cancer treatment. This is exemplified by Pfizer employing IBM Watson’s machine learning capabilities in their search for immune-oncology drugs and Genentech using GNS Healthcare’s AI system in their hunt for cancer treatments.
In recent years, there has been an exponential increase in the number of start-ups which use of AI to discover new drugs. The science blog BenchSci highlights at least 121 startups using AI for drug discovery. An example is Causaly, an AI platform which employs several AI technologies in tandem to analyse linguistic and statistical models from a plethora of scientific articles so as to decipher meaningful causal relationships and present them graphically. This enables them to answer complex questions such as: “What are the factors for liver cancer in the context of diabetes?” and “What pharmacological substances prevent depression?”
(2) Population health management – eg. Mental health
Given the increased awareness of mental health issues, the applications of AI in such interventions are becoming more relevant. One example is the conversational agent Woebotdeveloped by clinical psychologist Alison Darcy and her team at Stanford University. Woebot is trained with AI using the concepts of cognitive behavioural therapy (CBT). CBT is increasingly recognized as a form of mental health treatment which patients could learn to implement independently via a virtual guide (such as Woebot) as opposed to a human psychiatrist. This overcomes issues associated with long distance travelling, appointment scheduling and the stigma attached to mental health as patients may feel more comfortable conferring with a digital agent as opposed to a human being.
AI can also diagnose disease more accurately. Researchers at the UCLA Samueli School of Engineering have created a device which uses AI to diagnose parasitic infections by analysing bodily fluid samples to pinpoint moving parasites accurately and efficiently. The device is built with an algorithm for motion analysis which identifies the movement of a parasite and converts this into a visual signal which can be detected. This overcomes the time consuming challenge of manually searching for parasites in a drop of blood. Furthermore, the high sensitivity, low cost and portability of the device makes it appropriate for applications in developing regions where such infections are endemic.
The Babylon diagnosis and triage system is another AI initiative which is equipped with symptom checker functions to aid triage and diagnosis. This initiative supports diagnosis by depending on a probabilistic graphical model which draws on dependencies between symptoms, risk factors and medical histories.
Some menial yet important tasks such as ordering tests, writing up chart notes, and conveying medication prescriptions to the pharmacist, which are usually communicated between doctors and nurses could potentially be replaced by technologies facilitating voice to text transcriptions. Not only would this free up nurses’ time, it would also enable smoother healthcare delivery through increased efficiency and reduced likelihood of errors such as forgetfulness or oversight.
A partnership between IBM Watson and Cleveland clinic was established with the intention of developing an “Electronic medical record assistant” which would scan through and summarise patient histories, to provide useful insights on their needs and what the most appropriate treatment plan would be for them.
Some potential hurdles that researchers and bioentrepreneurs should take note of
The road to the implementation of any new technology is often fraught with obstacles. Likewise, although AI offers much promise in enhancing medicine and life sciences applications, there are several hurdles to overcome before researchers and bioentrepreneurs can harness the full potential of AI.
Inability to display genuine emotion or common sense
While AI has been lauded as a technology which may make humans obsolete, it is important to note that it has yet to surpass human intelligence in the ability to display genuine emotion or common sense. An AI chatbot Wysa, implemented to aid users tackle mental health conditions using conversation as a means of therapy, was condemned for its inability to react appropriately to victims who had shared their woes regarding sexual abuse. The AI agent, during the course of the conversation, failed to sense the magnitude of the issue and did not urge the user to seek immediate help. Human assessment is required in such instances so that the appropriate interventions can be taken. Hence, AI should be regarded as a complement to human labour, not a replacement!
A lack of technical expertise is an issue frequently flagged by researchers and resolving it tacitly requires collaborations with experts in AI and machine learning. However, these experts may not be specifically trained to apply AI in life sciences (eg. to analyse and interpret data from single cell sequencing, proteomics, metabolomics etc.). Computational biologists who are trained in AI technology and life sciences are key for such collaborations. Using their expertise as biologists and computer scientists, computational biologists have developed software such as CellProfiler, which uses deep learning to measure cell characteristics without human intervention and Deep variant, which also employs deep learning to provide thorough annotations of the genome.
High complexity of AI
Furthermore, the large number of knowledge gaps yet to be decrypted, makes it very challenging and potentially daunting for many scientists and healthcare researchers to incorporate AI in their innovations. John McCarthy, a prominent computer scientist in the AI scene, had initially suggested that language and other problems that remain elusive in the AI field could be “solved in a mere summer with the right group of scientists”, was forced to admit the opposite. However, as more research is conducted and more insight is gained into AI, the knowledge gaps should gradually decrease. While this may seem somewhat far-fetched at the moment, we have actually achieved plenty in just the last decade. One example is Apple’s digital assistant Siri. The image below shows two successive commands posed to Siri, the first being “Find restaurants near me”, which both versions of Siri correctly responded to. However, the 2014 version of Siri could not complete the request “Tell me more about the second one” when asked for suggestions about the second restaurant in the list. However, an updated version of Siri in 2017 responded with appropriate suggestions when asked to identify Italian restaurants from a list. This reiterates that we have gradually been overcoming the challenges posed by the complexity of AI and will continue to do so, potentially at a quicker pace now, as the knowledge on the topic grows.
Black box effect
The “black box effect” occurs when there is no clear explanation as to how the output was generated at the processing and deduction stages in between the output and the input (ie. the black box).
This black box effect currently appears to be an inevitable consequence of AI employment. As such, precautions need to be in place such that responsibility is taken or assigned, for the advice and actions suggested by the AI agent to the user. To address this issue further, you may choose to reduce the complexity of your intervention, to the point where you can gain a firm grasp of the parameters of the AI agent. With this level of control in place, the black box effect can be significantly diminished.
The machine learning and deep learning components of AI depend on the input data. What if that information is skewed? For example, consider an AI platform employed to diagnose people with Alzheimer’s disease by assessing their fluency of speech and ability to recall memories based on prompts. If the machine learning for this software was trained by getting native English speakers from the UK to interact with the system and it was then used to diagnose a native Spanish speaker who spoke English as a second language, their pauses, lack of fluency, differences in intonation and pronunciation may be misinterpreted as indicators of Alzheimer’s onset.
Hence, even if the data sets are not biased, they need to be very carefully crafted to prevent the inclusion of social bias. This is vital as AI algorithms tend to be “slaves to the data from which they learn.” The lack of common sense and dependence on input knowledge means that AI technologies do not take into consideration anything else besides what is provided to them, including an inability to detect bias in their recommendations and outputs.
Conclusion – final remarks and take-home messages
AI penetration in healthcare is gaining traction as evidenced by the burgeoning number of AI-related health start-ups. AI-based technologies have been posited to be involved in four broad areas in healthcare: therapeutics (eg. pharmagenomics to guide drug development),
population health management (eg. healthy lifestyle interventions), diagnostics (eg. deep learning to diagnose medical conditions from image analysis) and administration (eg. big data for hospital management). Future interventions are likely to be focused on advancing these applications which have shown potential for AI involvement.
To harness the full potential of AI, there are several issues that should be addressed. While the complexity and intricacy of AI has not been fully uncovered, the field is advancing rapidly and there is hope that bio entrepreneurs, aspiring to bring their research beyond the bench and into the community with tech savvy initiatives, won’t be kept in the dark for too long!
About the Author
Dhakshenya is a PhD student at LKC Medicine (NTU) and is currently in her second year. Her undergraduate degree was in biomedical sciences (Imperial College London) and her PhD explores the use of conversational agents for people at risk of diabetes in Singapore. During her graduate studies, she became acquainted with artificial intelligence through exploring the distinction between AI and rule-based healthcare chatbots and reviewing the literature evaluating these conversational agents. Besides research, Dhakshenya enjoys writing and is an avid reader. She has previously dabbled in article writing for lifestyle magazines but is making her scientific writing debut at Biotech Connection Singapore!