Group 3: Ethics 3 - Should machines diagnose patients?

Introduction

We are, for the first time in the history of this universe connected to each other like never before. As per International telecommunication union 40% of the humans on this planet are connected to the world wide web. This connected system generates 2.5 Exabytes of data per day (Mikal Khoso. et.al). We have been successful in developing technologies that can generate insights using this data. Insights ranging from weather stats to finding out floating rumors on the social media thus avoiding a social problem, are a result of these intelligent technologies. These technologies, till date have never failed us, but only helped us deepen our understanding of the nature. 

Below we discuss some problems and opportunities in using machines and the generated data in the hospitals for patient diagnosis.

The case of Hospitals:

Hospitals generate a lot of data per day. Although this data is not in a structured form but in the form of CT-scans, MRI scans, X-rays, Prescriptions, Textual reports etc. we can use this data to learn about the patient. Reading from data has never been easy for a human, Humans are not designed to derive insights from large sets of data. In such circumstance, using machines to derive useful insights about the patients, can enhance the doctors abilities to perform their tasks more precisely. 

The scale problem:

With the increasing population, the hospitals are also faced with the problem of scale. The rate of population growth is not in scale to the rate of induction of new doctors in the system. Hence, a doctor has to handle large number of patients in the same available time. This results in allocation less time per patient or large waiting times. Large waiting times increase the risks in patient health. Less time per patient risk the possibility of the patient being diagnosed incorrectly.  In such situations, use of machines in the early stages or diagnostic stages can be helpful for the doctor as he might get a idea about all the patient vitals to make informed decisions about the treatment to suggest.

IBM Watson and other AI

IBM Watson and Watson Health:

Watson is a question answering (QA) computing system that IBM built to apply advanced natural language processing, information retrieval, knowledge representation, automated reasoning, and machine learning technologies to the field of open domain question answering. (Source :  "DeepQA Project: FAQ". IBM. Retrieved February 11, 2011.) 

Watson Health is a product offering by IBM that is designed to become clinical decision support systems for the use by medical professionals. As every intelligent machine needs training, Watson health relies on data sources like treatment guidelines, electronic medical record data, notes from physicians and nurses, research materials, clinical studies, journal articles, and patient information (Source: "Putting Watson to Work: Watson in Healthcare". IBM. Retrieved November 11, 2013.)

Below are some of the case studies when Watson was used in diagnosis

Watson Health in Brain cancer:

IBM is working on development of the Watson system towards transforming care for all types of cancer, based on the genetic characteristics of that person’s cancer. The work towards development of the product involves development of a solution specifically targeted at interpreting data from genomic data. By analyzing variations in normal and cancerous biopsies of brain tumors, Watson can then suggest treatment options that would be tailored to an individual’s specific type and personalized instance of the cancer. Watson will do this with the help of medical literature and clinical records as data sources.

The Watson system is designed to complement rapid genome sequencing and to help dramatically reduce the time from gathering the genomic data of an individual's tumor variant to clinical interpretation (Royyuru, Ajay et.al) . This will enable clinicians to more rapidly make decisions on how they can treat their patients. The reduction in time spent from genetic analysis to treatment planning can be very helpful in cases of glioblastoma which has a median  survival of less than 15 months.

In a recent study Watson for Genomics was able to provide whole genome sequencing (WGS) data – complete with clinically actionable insights – from the tumor cells within 10 minutes. The traditional technique took around 160 hours of manual human analysis to arrive at similar conclusions (Ricci, Marco.et.al).

Should we store all of our medical records in the cloud?

Clouds provide us with a lot of advantages which include,

1. Access to data with geographical transparency.
2. Guarantee of data retention in case of natural disasters
3. Protection towards accidental deletion of data. and many more
4. Cost saving

With these advantages comes the fear of data ownership and privacy. Medical records especially being extremely sensitive might suffer from extreme negation towards cloud storage. However there are a variety of ways how these problems are getting solved.

1. Use of Block chain architectures to store Patient records
2.  Encryption techniques
3. Use of personal clouds to store data
   
   

Can a doctor rely on AI to gather all information necessary to diagnose a patient?

Artificial intelligent systems can be efficient in providing a diagnosis based on the previous sample cases it has seen. However the major drawback of the AI systems currently is their inability to provide explanations for the results. A deep learning system is typically a black box with a input and output port. We pass in the input sample and generate a result, without really knowing the basis of the result. We trust the system because the output that it generates matches the expected results. Say, when we input a image of a dog to the system, it classifies it as a dog and not a cat, which is visually verifiable. This may not be the case in cancer diagnosis. 

This nature of the AI system can be a cause of concern for the doctors when using AI for diagnosis of a patient.

Deep Learning for diagnosis

(Mukherjee, Siddhartha)  talks about 'computer aided diagnostic'  in mammography where a rule based software is capable of highlighting suspicious areas for the radiologists to review. These systems, being rule based are programmed with a set of rules that help them make diagnostics. A program that has seen 3 X rays can make the same diagnostics as the program that has seen 3000 X rays. As the rules are static there is no dynamic learning involved.  A neural network on the other hand is inspired by the model of how the brain works. In the brain the neural synapses get stronger or weaker based on the use of those neurons. The neural networks mimic the similar behavior of the neurons using mathematical functions.

(Mukherjee, Siddhartha.et.al) explains how the neural networks work by using a simple example :

“Imagine an old-fashioned program to identify a dog,” he said. “A software engineer would write a thousand if-then-else statements: if it has ears, and a snout, and has hair, and is not a rat . . . and so forth, ad infinitum. But that’s not how a child learns to identify a dog, of course. At first, she learns by seeing dogs and being told that they are dogs. She makes mistakes, and corrects herself. She thinks that a wolf is a dog—but is told that it belongs to an altogether different category. And so she shifts her understanding bit by bit: this is ‘dog,’ that is ‘wolf.’ The machine-learning algorithm, like the child, pulls information from a training set that has been classified. Here’s a dog, and here’s not a dog. It then extracts features from one set versus another. And, by testing itself against hundreds and thousands of classified images, it begins to create its own way to recognize a dog—again, the way a child does.” It just knows how to do it.

Tumor classification case study (Beckett, Jamie et.al):

The aim of this case study was to detect the presence of a tumor and classify it into one of several broad categories, using deep learning. More than 100 brain tissue samples are used in the case study as input. The deep learning algorithm analyses tissue images from Raman histology. Doctors currently halt during the surgery for around 30 to 40 minutes for the tissue samples to be processed in the labs. The Raman histology reduces this wait time to 3 minutes. This makes the diagnosis of the tumor possible in the operating room itself. 

The deep learning algorithm classifies the tumors in four categories. The accuracy rates of the algorithm are 90 percent which are comparable to that  of the neuropathologists, 95 percent . Deep learning and the SRH imaging technique can thus help the doctors to make better decisions and decide whether to operate or not. 

Legal Issues

Who is responsible if a physician relies on a false prediction by a computer program?

Intelligent machines are trained from data and other signals such as emotions are not considered by the machine. So, the times when the machine makes a mistake can be considered as a logical flaw in the system. Inconsistencies in the system can be a result of,

1. Incorrect models used to design the algorithms.
2. Improper data sources used while training the algorithms.
3. A bug in the algorithm. 

In order to get rid of all these error possibilities or inconsistencies, it is extremely essential for the machines to be used for diagnosis, at least in the initial phases (During the testing phase). This will ensure that the error detection process starts early. Once the results generated by the machines are in sync to the ones generated by doctors, then we can be certain about the performance of the machines. This will ensure that the cost of false predictions remains low initially. 

Author of a Software is responsible for the inconsistent behavior of the system. It is the responsibility of the author to publish all the known issues in the software before it is released. Intelligent machines are software systems. In this case, any error or inconsistently should be owned by the author of the system. 

Is it legal in Europe to rely on artificial intelligence or computer prediction?

A briefing on AI: potential benefits and ethical concerns, at the EU parliament talks about concerns like, data privacy, ownership, algorithm transparency and accountability. The main point being, trusting the AI systems. It is of extreme importance that we trust AI systems and make sure that they follow the same ethical principles, moral values, professional codes, and social norms that we humans would follow in the same scenario.

Trust can be built by accountability and hence the algorithms that are used by the AI systems should be transparent. The AI algorithm should be able to explain their behavior while making a decision in a way which can be understood by a human. Getting a understanding of how the systems work can increase the trust towards the AI systems. To make this possible, a AI system should be a explanation-based collateral systems. In addition to this, Ethics module should be constantly adapted to reflect the best practices (Rossi ,Francesca et.al).

Final decision or can doctors be replaced?

To reap the societal benefits of AI systems, we will first need to trust it. The right level of trust will be earned through repeated experience, in the same way we learn to trust that an ATM will register a deposit, or that an automobile will stop when the brake is applied (Rossi, Francesca et al.). A human starts trusting things when he is able to trust a system. At this stage the major roadblock in the use of AI for medical diagnosis is the inability of the AI to generate explanations. Without the right level of details on the decision process, it would be unwise to trust the systems at this point in time.  Research efforts towards development of AI systems that have good explanations built in the algorithms can help make the AI usable in medical diagnosis. 

However as mentioned earlier, there are some tasks, even at this point of time which do not need a explanations. Automating these tasks such as, finding patterns in the scans etc, can reduce significant time in the treatment. 

Bibliography

International Telecommunication Union (ITU), World Bank, and United Nations Population Division, Internet live stats. [Online] Available at : http://www.internetlivestats.com/internet-users/

Mikal Khoso .How much data is produced everyday.  [Online] Available at : http://www.northeastern.edu/levelblog/2016/05/13/how-much-data-produced-every-day/

Christopher Chabris, Daniel Simons. The invisible gorilla.  [Online] Available at : http://www.theinvisiblegorilla.com/gorilla_experiment.html

Alix Spigel. Why Even Radiologists Can Miss A Gorilla Hiding In Plain Sight .[Online] Available at : http://www.npr.org/sections/health-shots/2013/02/11/171409656/why-even-radiologists-can-miss-a-gorilla-hiding-in-plain-sight

Royyuru, Ajay, IBM Watson takes on Brain cancer. [Online] Available at : http://www.research.ibm.com/articles/genomics.shtml

Ricci, Marco. IBM Watson impresses in brain tumor analysis. [Online] Available at : https://pharmaphorum.com/news/watson-for-genomics-brain-tumour/#

Mukherjee, Siddhartha .AI vs MD [Online] Available at : http://www.newyorker.com/magazine/2017/04/03/ai-versus-md

Beckett, Jamie. Brain Trust : How AI is helping surgeons improve tumor diagnosis [Online] Available at : https://blogs.nvidia.com/blog/2017/02/17/ai-helps-improve-brain-tumor-diagnosis/

Rossi, Francesca. AI: Potential benefits and ethical considerations. [Online] Available at : http://www.europarl.europa.eu/RegData/etudes/BRIE/2016/571380/IPOL_BRI%282016%29571380_EN.pdf