Artificial Intelligence for Multiple Long-Term Conditions (AIM) - - PowerPoint PPT Presentation

Artificial Intelligence for Multiple Long-Term Conditions (AIM) Engagement Workshop 2

17th July 2020, 10am-12pm

Where in the world are you today? Pin your location to the map! (link in the chat)

https://padlet.com/aimcallmltcs/8uydjkku6hpbjy5r

Welcome and Agenda:

1000 Welcome, introduction to the call and reflections on the first engagement event 1010 Introduction to the NIHR Research Design Service (RDS) 1020 Keynote speakers:

• Professor Ben MacArthur, Deputy Programme Director for Health and Medical

Sciences, Alan Turing Institute

• Professor Tom Marshall, University of Birmingham, NIHR-MRC 2018

Understanding multimorbidity in the UK call

• Joyce Fox and Kate Ripley, Experts by experience

1045 Breakout session 1: 60 second pitches and discussion (Zoom) 1115 Coffee Break 1125 Breakout session 2: Choose your breakout session (MS Teams)

• Patient and public involvement
• Learning from failure
• NIHR Research Design Service
• Data

1145 Application process, future events and Q&A

Padlet: Locations of Participants

Overview:

Use AI and data science methods, combined with existing methodology and expertise in clinical practice, applied health and care research and social science, to systematically identify, map or explore clusters of disease Seek research to better understand the trajectories of patients with MLTC-M over time and throughout the life course, including the influence of wider determinants such as environmental, behavioural and psychosocial factors To fund multidisciplinary Research Collaborations to undertake programmes of work:

• Development Awards of up to £120k over 8 months
• Research Collaborations of £2.5-5m over 36 months (wave 1) or 30 months

(wave 2) Establish a Research Support Facility (RSF) (£3m) to support successful applicants.

• Ben MacArthur: Professor in Mathematics at the Life Science

Interface, University of Southampton and Deputy Programme Director for Health and Medical Sciences, Alan Turing Institute.

• Tom Marshall: Professor of Public Health and Primary Care, University
• f Birmingham. Funding award holder for NIHR-MRC project - Bringing

Innovative Research Methods to Clustering Analysis of Multimorbidity (BIRM-CAM)

• Joyce Fox and Kate Ripley: Experts by experience and members of

the Economics of Health and Social Care Interface Policy Research Unit PPI Group.

Keynote Speakers:

NIHR AIM Call RDS Support & Advice

Jörg Huber (RDS SE), Jane Fearnside (YH) & Bernadette Egan (SE)

www.rds-se.nihr.ac.uk Twitter: @NIHR_RDS 17th July 2020

• FREE confidential

support for health and social care researchers across England on all aspects and methods of research design and grant application development

• Expert RDS advisers can

help with all aspects of designing a proposal incl.:

• research design and methods
• funding sources
• refining research question
• outcome measures
• PPI and building a team
• avoiding common pitfalls

About RDS

We offer support along the way to ‘pressing the button’

Key elements central to NIHR applications

• Project
• People/team
• Places
• Public involvement (PPI)

Priorities for NIHR:

• Diversity
• Inequalities and disadvantage

Read and follow the guidance. If and when interacting with us, be prepared for us not to agree with you.

Resources

• https://www.nihr.ac.uk/documents/research-on-multiple-long-

term-conditions-multimorbidity-mltc-m/24639

• https://acmedsci.ac.uk/file-download/99630838
• https://bjgp.org/content/68/669/e245
• Kastner M, Hayden L, Wong G, et al. Underlying mechanisms of

complex interventions addressing the care of older adults with multimorbidity: a realist review. BMJ Open 2019;9:e025009. doi:10.1136/bmjopen-2018-025009 Contact your RDS

• https://www.nihr.ac.uk/explore-nihr/support/research-design-

service.htm

• https://twitter.com/nihr_rds?lang=en

Thank you Join RDS Teams Meeting later on +44 20 3443 8728 United Kingdom, London (Toll) Conference ID: 402 821 178# See Agenda for today

Finding expert partners

• On multi-morbidity
• Geriatrician
• Generalist physicians
• Core problem is shortage of doctors
• RDS can help with this through putting out calls

across the regions Contact your RDS

• https://www.nihr.ac.uk/explore-nihr/support/research-

design-service.htm

Health and Medical Sciences

• Prof. Ben MacArthur

NIHR AIM call: Opportunities for data science

• This multi-disciplinary / cross-disciplinary / cross-institutional call
• ffers tremendous opportunity to develop new ways of approaching

multimorbidity.

• New way of conducting research that develops a holistic view of life

course health that combines physiological, psychosocial and environmental factors, and learns from heterogeneous linked data.

• Better understanding of disease clusters and trajectories will be

needed to develop integrated treatment approaches to meet the needs

• f people with MLTC-M.
• This is hugely challenging and will require new ways of working that

have not been developed yet (including ethics and data security).

NIHR AIM call: Challenges for data science

• To meet these challenges requires new analysis tools that can handle

complex, distributed data.

• Innovation is needed in areas such as: model interpretability, causal

inference, missing data, combining mechanistic and statistical models, machine learning, uncertainty quantification …

• Bring together mathematicians statisticians, computer scientists toward

solving this grand challenge.

• This challenge = a great opportunity to think creatively about new ways
• f doing multi-disciplinary data science.

NIHR AIM call: Summary

• Vital that these approaches are developed in close collaboration with

clinical, health care research expertise and have clear benefits to patients.

• Requires careful consideration of how analysis will reduce health

inequalities.

• Vital that we collectively develop a coordinated approach: call will

stimulate collaboration between groups, assisted by a central facility.

• See NIHR Strategic Framework for MLTC-M Research and the cross

funder AAS, MRC, NIHR, Wellcome multimorbidity research framework.

Bringing Innovative Research Methods to Cluster Analysis of Multimorbidity (BIRMCAM)

Tom Marshall

1

Collaboration

• University of Birmingham
• Institute of Applied Health Research
• University of Cambridge
• MRC Biostatistics Unit
• Department of Public Health and Primary Care

2

Aims

• Develop understanding of multimorbidity clustering methods
• Objectives:
• Critical review
• Create a “methodological commons” of multimorbidity clustering methods
• Apply new techniques: probabilistic machine learning (ML); multistate models

that predict transition / trajectory

• Implement and validate findings in two large primary care databases

3

So far

• Identified 5 main approaches to multimorbidity clustering
• Latent class analysis (LCA)
• Hierarchical cluster analysis (HCA)
• Multiple Correspondence Analysis followed by k-means (MCA-k)
• k-modes (kmodes)
• For binary data
• k-means (kmeans)
• Uses Euclidean distance & generally used for continuous variables

4

Simulated datasets

• Clustering patients (not diseases)
• Use Rand index to compare method to known number of clusters

5

Simulated dataset characteristics

1. Prevalence

• Frequency of underlying condition

2. Noise

• Random error

3. Number of clusters 4. Frequency of ‘nulls’

• Number of individuals not in any clusters

5. Correlation

• Between diseases within a cluster

6. Overlap

• Diseases can belong in multiple patient clusters

7. Balance

• Vary number of patients in each cluster

6

Broad findings

• Rand index declines with
• More noise
• Lower prevalence
• Some methods tend to perform better
• LCA (Latent Class Analysis)
• HCA (Hierarchical cluster analysis)

7

Primary Care Records Analysis

• Simplified primary care dataset
• Real patients
• Limited number of morbidities

8

Characteristics of dataset

9

Methods

• THIN (IMRD) dataset
• age 65-84y, n=6,387 patients with ≥1 condition
• excluding conditions with prevalence <2%
• Bootstrap 500 samples from original dataset
• Pre-specify the number of classes (clusters)
• Run analyses
• Summarise, for most frequently occurring clusters

i. Proportion of dataset in cluster (median %, IQR) ii. Frequency of occurrence of each cluster (as % of 500 bootstrap samples) iii. Three most frequently occurring conditions in cluster (median %, IQR)

10

Age 65-84y: Latent Class Analysis (5 classes)

• 1. Affective psychosis cluster - 6.4% (5.4 – 7.1) 417/500 (83%)

Affective 100% Fractures 35.8% (32.8 – 39.9) Non-affective 33.7% (31.1 – 38.1) 234/500 (47%) Affective 100% Fractures 37.2% (35.4 – 39.4) Pneumonia 31.6% (29.8 – 33.3) 92/500 (18%) Affective 100% Fractures 31.7% (34.5 – 39.1) UTI 29.4% (27.1 – 31.5) 31/500 (6%) Affective 95.7% (91.3 – 100) Non-affective 52.1% (38.1 – 63.7) UTI 32.5% (29.6 – 34.5) 28/500 (6%)

• 2. Fractures cluster - 24.9% (21.8 – 27.4) 496/500 (99%)

Fractures 100% Pneumonia 49.0% (46.1 - 51.4) UTI 34.4% (32.2 – 36.8) 420/500 (84%)

• 3. Depression cluster - 25.9% (23.0 – 27.8) 500/500 (100%)

Depression 100% Pneumonia 39.3% (34.5 – 41.2) Fractures 34.8% (33.2 – 36.5) 345/500 (69%) Depression 100% Pneumonia 38.9% (36.3 – 42.6) UTI 32.0% (30.2 – 33.3) 90/500 (18%) Depression 100% Pneumonia 34.8% (32.6 – 37.3) Affective 31.3% (29.3 – 34.0) 47/500 (9%)

• 4. Diabetes cluster - 17.1% (15.7 – 18.0) 353/500 (71%)

Diabetes 100% Pneumonia 35.4% (33.2 – 37.0) Fractures 31.6% (30.2 – 33.0) 269/500 (54%) Diabetes 59.9% (58.0 – 64.0) Stroke 49.6% (45.5 – 51.7) MI 41.4% (39.6 – 43.4) 49/500 (10%)

11

• 5. Pneumonia cluster – 26.4% (21.3 – 28.3) 500/500 (100%)

Pneumonia 75.2% (74.2 – 76.3) UTI 54.5% (53.2 – 55.7) Stroke 29.9% (28.5 – 31.1) 270/500 (54%) Pneumonia 86.5% (84.8 – 88.4) UTI 71.2% (68.5 – 73.1) Flu 31.1% (28.4 – 33.8) 58/500 (12%) Pneumonia 68.7% (53.4 – 100) Diabetes 39.2% (28.0 – 45.3) Stroke 37.2% (30.9 – 46.6) 95/500 (19%)

• 6. UTI cluster – 19.7% (14.9 – 24.9) 155/500 (31%)

UTI 100% Pneumonia 54.7% (52.2 – 56.3) Fractures 32.6% (31.6 – 34.3) 45/500 (9%) UTI 100% Pneumonia 61.8% (60.5 – 62.9) Diabetes 20.5% (19.7 – 21.7) 42/500 (8%) UTI 52.9% (51.7 – 54.7) Diabetes 46.6% (46.0 – 48.2) Stroke 43.0% (41.8 – 44.1) 29/500 (6%)

12

Age 65-84y: Latent Class Analysis (5 classes)

Age 65-84y: MCA-kmeans (5 classes)

• 1. Fractures - 17.9% (17.0- 19.5) 87/500

Fractures 59.0% (56.3 – 60.5) Diabetes 53.8% (51.8 – 56.7) Stroke 48.0 % (45.1 – 51.3) 76/500 (15%)

• 2. Depression - 14.9% (10.1 – 19.4) 1070 (multiple clusters within sample)

Depression 89.7% (88.3 – 91.6) Affective psychosis 59.7% (57.5 – 61.8) SMI 47.4% (45.0 – 50.9) 456/500 (91%) Depression 81.1% (79.6 – 83.1) Pneumonia 45.5% 943.6 – 47.9) Fractures 39.9% (38.2 – 41.8) 273/500 (55%) Depression 88.6% (86.8 – 92.4) Pneumonia 65.0% (54.6 – 69.0) UTI 46.7% (42.3 – 50.7) 133/500 (27%) Depression 65.1% (62.9 – 69.5) Fractures 48.7% (46.3 – 51.2) Affective psychosis 42.4% (38.9 – 48.6) 151/500 (30%)

• 3. Diabetes – 17.6% (16.3 – 18.8) 383/500

Diabetes 59.6% (56.8 – 62.0) Fractures 50.3% (47.6 – 53.5) Stroke 50.0% (47.5 – 52.2) 342/500 (68%)

• 4. Pneumonia – 29.4% (21.3 – 31.0) 918 (multiple clusters within sample)

Pneumonia 60.9% (59.6 – 63.2) Fractures 46.7% (45.6 – 47.8) UTI 38.8% (37.6 – 40.7) 487/500 (97%) Pneumonia 85.8% (84.7 – 86.7) UTI 71.7% (69.8 – 73.2) Flu 34.7% (33.2 – 36.8) 326/500 (65%) Pneumonia 80.6% (79.0 – 85.5) UTI 62.1% (60.7 – 63.8) Depression 40.7% (36.6 – 42.7) 83/500 (17%)

13

Similar analyses for other methods

• Same method repeated does not always identify the same clusters
• Latent Class Analysis appears to perform better
• Limitations to unsupervised clustering methods

14

Reflections on PPI in research, Kate Ripley

Patients as partners in research bring greater focus and relevance for the people most directly affected by health and social care developments. Empowering the Public and Patients:

• Provide opportunities to help identify and frame research questions
• Provide opportunities to verbalise what the proposed research means

for people (i.e. their lived experience) and how it will influence research

• Draw upon being part of something bigger; the wider objectives and

mission of the NIHR around PPI

Providing Learning Opportunities

• Contributions to PPI Engagement Strategy
• Promoting PPI at events, encouraging others to get involved
• Completion of online learning modules provided by NIHR

(Intro to NIHR & PPI in research & How to review research documents from a public & patient point of view)

• There are links to many other sources of online learning and

YouTube videos

How to involve the public in research

• Make language accessible & simple - remember the average UK reading age
• Appeal to people’s interests/emotional needs capitalising on how they could

influence the direction of research (people usually are interested if it relates to their health or that of a relative)

• Tap into health charity promotional channels (Alzheimer's Society)such as e-

newsletters to gain interest and promote NIHR at the same time

• Offer opportunities in a variety of ways not just online as this excludes a large

proportion of people either due to poverty and/or lack of expertise

• Never underestimate word of mouth as a communication channel
• Don’t be too wordy or technical (automatic switch off)
• Be personable and approachable (some people can feel intimidated by

academics – Imposter Syndrome)

Making Data less ‘Scary’

• Watch out for language- most lay people don’t know what mortality
• r morbidity are –they know illness and death
• Encourage others to look at the narrative behind statistics making it

relevant & real The NIHR expects the active involvement of patients and the public (e.g. service users and carers) in the research and infrastructure that it supports, where appropriate as per INVOLVE guidance.

Breakout 1: Networking Pitches

You have 60 seconds to pitch the following information to the rest of the group:

• Who are you and where are you?
• What is your area of expertise?
• How is this is relevant to the AIM call?
• What is the ‘missing link?’ i.e. What

disciplines/expertise/data you need to take your research to the next level?

Followed by discussion.

Coffee Break (10 mins)

Network with other participants and post on our expertise noticeboard (link in the chat)

https://padlet.com/AIMCallMLTCs/beeagk8njhavrl9b

Breakout 2:

Choose which of the breakout session discussions you would like to attend on MS Teams. You will need to leave the main Zoom meeting to join the breakout session, before rejoining the main Zoom meeting (links are provided in the chat and on your agenda)

• Patient and public involvement
• Learning from failure
• NIHR Research & Design Service
• Data

Application Process, Future events and Q&A: Mario Moroso, Assistant Director, Research Programmes, NIHR