Artificial Intelligence in Healthcare: Regulatory Compliance, Human Activity Recognition, Genomics,

Regulatory Compliance for Machine Learning Systems in Healthcare

Question 1

Dr. Ali, a colleague in the cardiology ward at the UK NHS hospital you work in, is unfamiliar with artificial intelligence. They have hired a specialist company called CleverAI Ltd to develop a machine learning system to predict the length of a patient’s hospital stay (i.e. number of days from admission to discharge). This information may be used to plan staffing levels in the department or to adapt the care plans for patients who are expected to be in the ward for longer. Dr. Ali has suggested that the input data for this model could include the following:

Patient’s clinical history (prior diagnoses, medications, etc)
Patient’s current treatment plan (surgeries, medications, and other forms of treatment/care)
Other patient information (date-of-birth, height, weight, gender, marital status, ethnicity, home address, date of admission and discharge for any previous hospital visits)
Hourly ward data for the past 10 years (number of staff on duty, number of beds occupied/available)
External data, captured daily (local weather, prevalence of major infectious diseases in the region)

In partnership with the hospital, Dr Ali plans to establish a spin-out company so that if the project is successful, the system can be licensed for use in hospitals in the UK and around the world.

a) CleverAI Ltd have not previously worked on healthcare applications, so Dr Ali wants your help to ensure the project follows an appropriate process to achieve regulatory compliance. Outline the key steps in the pathway to NHS regulatory compliance for wider use. Briefly explain (1-3 sentences) for Dr Ali what each step means in the context of this particular project.

b)In one of their sprint reports for Dr. Ali, CleverAI included the following graph (Figure 1) without thorough explanation. All they said was that it showed results from their initial attempts at training a model for this project.

Question 2 (25 marks)

Human activity recognition (HAR) aims to recognise activities from a series of observations on the actions of subjects and the environmental conditions. Its applications can be seen in hospitals and clinics, which often provide a continuous and non-invasive awareness of activity that can provide clinicians, nurses, and other healthcare workers with assistance such as patient monitoring, automated documentation, and monitoring for protocol compliance. 

Propose an HAR system for the above-mentioned purpose (i.e., monitoring of activities in hospitals and clinics) and answer the following questions:

a) What activities could be monitored and for what purpose? Why is computer vision deemed a suitable solution?(4 marks)
 
b) What visual sensor(s) do you need? How do you set up the hardware system?(4 marks)
 
c) What image processing/computer vision/machine learning algorithms can you employ to enable this application?(6 marks)
 
d) What is the process of evaluating the performance of the proposed system?(5 marks)
 
e) What are the economic and wider societal benefits of this application? Are there any issues surrounding the introduction of this technology in its commercial, socio-economic, environmental and cyber security context?

Question 3 (25 Marks)
 
a) Genomics: Provide an example where you would use target sequencing. Name a target sequencing pipeline/workflow and outline its main steps.(4 marks)

Human Activity Recognition for Patient Monitoring in Hospitals and Clinics

b) DW Health is a private healthcare company, and is considering adopting the use of Electronic Health Records across their services. DW Health operates in the UK, and their services often integrate with services delivered by the National Health Service (NHS), who will need to access the information in the EHRs. They have come to you for advice on how this decision would affect their patients.

1) Describe clearly one key benefit to their customers (patients) and one key concern patients might have about use of EHRs by their healthcare provider.
(4 marks)

2) Much of the key information in the company’s existing records is stored as typed comments from clinicians, stored in text files. State the name of the general category of AI that might allow key information from these written notes to be automatically identified and extracted.
(2 marks)

3) To ensure secure and reliable transfer of sensitive data from these EHRs to NHS hospitals and other service providers when necessary, DW Health is planning to develop a new application programming interface (API). Is this a good idea? Explain in detail why or why not.
(5 marks)

4) Describe two other technical challenges (not those in the parts of this question above) associated with the application of artificial intelligence to Electronic Health Records and describe, for each challenge, how it can be addressed. (8 marks)

c)    Write an SQL command to create a new database called Critical Care Episodes.

Question 4

Dr. UWE is the head of the radiography department in a UK hospital, and has come to you for advice because she is not very familiar with AI. She has been approached by LookSmart, an AI start-up company who have developed software to automatically detect tumours and other abnormalities in x-ray images, and she is trying to decide whether this tool could be helpful for her department. 

Dr. Uwe tells you “They explained to me that their algorithm is some kind of neural network that has learned particular patterns that occur in cases where a human expert has identified an abnormality. Effectively, it examines the image pixels within small rectangular windows, and slides these windows across the image to determine whether those patterns occur anywhere.”

a) Based on Dr. Uwe’s description, what general category of neural network do you think they are using? (2 marks)

b) Which of the following terms also describe this system? Write the numbers of all that apply.
1. Machine learning
2. Expert systems
3. Supervised learning
4. Unsupervised learning

c) LookSmart state that “Our AI is highly accurate. It was trained and evaluated using data from every major hospital in Brazil, including over 100,000 patients.” Dr. Uwe is wondering whether it’s good enough to use in her department. Based on this statement from LookSmart, give one reason to justify trust in the system and one reason to doubt the system’s trustworthiness.

d) Describe a specific example of bias that could hypothetically arise from LookSmart’s system based on the description given. Make sure you identify who could be affected and how. Also, briefly describe an approach LookSmart could take to reduce this bias in the system.

e) In accordance with good machine learning practice, to train and evaluate their system LookSmart split the Brazilian dataset into three subsets at random: training, validation, and test. The accuracy achieved in each dataset was as follows:
Training – 98 %
Validation – 83 %
Test – 81 %

LookSmart’s representative argues that, of the three, the Training accuracy of 98 % gives the most reliable representation of the system’s performance because the Validation and Test datasets were much smaller (roughly 10,000 patients each, versus 80,000 patients in the Training set). Which of the three accuracy figures do you think Dr. Uwe should pay the most attention to, and why?

f) Dr. Uwe would like to better understand the cybersecurity risks that might be associated with the use of LookSmart’s system. You explain to her that the ‘CIA Triangle’ is a useful framework for considering the range of security threats. For each of the three points in the triangle, describe a hypothetical example of a specific threat that might arise in the use of LookSmart’s system.