Advanced Computation Laboratory

1
Advanced Computation Laboratory

• Quinquennial review 2004

2
The Advanced Computation Lab.

• Like other labs in the LRI the ACL carries out
  basic research that contributes to the mission of
  CRUK in the understanding, prevention, diagnosis
  and treatment of cancer.
• Our basic research interests are in the
  computational and cognitive sciences.

3
Focus and methods

• The scientific focus is in understanding
  cognitive processes (e.g. reasoning,
  decision-making and planning, natural language)
  and their integration in natural or artificial
  agents
• Research methods include formal analysis and
  computational modelling, in silico simulation,
  and experimental and clinical studies.

4
Theory
Computational experiments
Trials
PROforma language
Engineered technologies
Applications
5
A diverse range of projects

• We hope to bring out three central themes over
  the presentations
• Practical contributions to cancer care
• A theoretical centre of gravity
• A range of technical opportunities arising from
  the theory
• And welcome comments on strategy as well as
  project specifics

6
Topics

• Motivation improving care
• Modelling and rational design of clinical
  processes
• OpenClinical
• The CREDO trial

• Principles of cognitive systems
• Natural language interfaces
• From standard pathways to individual care plans
• Modelling the cognitive organisation

7
Motivation improving care
8
Medicine is a humanly impossible task

• Up to 97,000 unnecessary deaths p.a. in the US
  are due to medical error. Total national costs
  between 17B and 29B
• Institute of Medicine, To Err is Human 1999
• In NHS hospitals overall rate of preventable
  adverse events of 11.7. A third of which led
  to disability or death, others are frequent,
  minor events but together have massive economic
  consequences
• Vincent et al, BMJ 2001

9
and in oncology

• Perhaps 16,000 lives could be saved if all
  current knowledge of cancer were properly
  applied
• ICRF Vision for Cancer, 1995
• There have been undoubted improvements in
  service delivery but there is still a sense that
  progress has been patchy and that much has yet to
  be achieved.
• National Service Framework Assessment of NHS
  Cancer Care 2001.

10
Promoting best practice
11
but

•   Busy clinicians have little time to read
•   Even if there is time, memories are
  unreliable, working pressures acute
•   Conventional guidelines address general
  principles of care, not the needs of individual
  patients

12
Can informatics help?
13
ALL Dose adjustment studyJ Bury, C Hurt, A Roy,
L Cheesman, M Bradburn, S Cross, J Fox, V Saha
(submitted)

• Objectives
• To assess the clinical value of a decision
  support system designed to assist with dosage
  adjustments during maintenance therapy for
  childhood Acute Lymphoblastic Leukaemia.
• Decision model
• one PROforma decision task, 8 options. Each has
  between 1 and 5 criteria associated with it, each
  referring to different clinical situations,
  expressed in terms of 5 parameters.
• Materials and methods
• Balanced-block crossover experiment, in which 36
  clinicians with varying degrees of experience
  were asked to decide on oral chemotherapy dosages
  for 8 simulated cases 4 using decision support
  and 4 without.
• Outcome measures were number of protocol
  consistent dosage decisions time to manage each
  case accuracy of dosage calculations and
  clinicians' opinions about the value and
  usability of the system.

14
Dose adjustment in chemotherapywith CRUK
Paediatric Oncology Group, London Hospital
Bury, Hurt et al, Proc. American Medical
Informatics Association, 2002 Hurt et al, Proc.
AI in Medicine Europe, 2003 Bury et al, British
Journal of Haematology (in submission)
15
ALL Dose adjustment study resultsJ Bury, C
Hurt, A Roy, L Cheesman, M Bradburn, S Cross, J
Fox, V Saha (submitted)

35/36 subjects said they would be likely to use
the system if it were available
16
Triple Assessment C Hurt, V Patkar, R Steele, T
Rose, M Williams, J Fox (report in preparation)

• Objective
• To evaluate the potential effect of
  PROforma decision support on clinical decision
  making with respect to national guidelines for
  Triple Assessment.
• Decision model
• Pathway included 4 decisions (familial
  risk, type of imaging, type of biopsy, and
  management).
• Materials and methods
• 15 hypothetical paper cases covering range of
  clinical scenarios developed by an expert panel
  of five judges (2 breast surgeons, 1 breast
  pathologist, 1 radiologist and 1 geneticist) and
  optimal management for each case was agreed by
  consensus.
• 24 doctors asked to manage 5 cases with and 5
  without computer support. A balanced block
  design used to allocate cases.

17
Triple Assessment Study Results C Hurt, V
Patkar, R Steele, T Rose, M Williams, J Fox
(report in preparation)
24 participants (17 consultants, 5 specialist
registrars, 1 nurse practitioner) Average number
of years in speciality 9.3 (range 1 - 23)
Would patient care improve with TA decision
support? In favour 16 Undecided 1 Against
7 Would they personally wish to work with TA
decision support? In favour 12 Undecided
3 Against 9
18
Topics

• Motivation improving care
• Modelling and rational design of clinical
  processes
• OpenClinical
• The CREDO trial

• Principles of cognitive systems
• Natural language interfaces
• From standard pathways to individual care plans
• Modelling the cognitive organisation

19
Modelling and rational design of clinical
processes
20
The domino model
Clinical goals
Patient data
Actions orders
Protocols pathways
Clinical options
Arguments
21
Decision-makingDose adjustment in ALL
Current dose, duration Platelets, Tolerance
Review dose
Leave dose unchanged
ALL97 Trial criteria
Increase, decrease, leave unchanged
22
Plans and pathwaystriple assessment (imaging
fragment)
Pain, discharge, nodule??
Report
Investigate symptoms for possible Ca

• Mammogram
• Ultrasound
• Registration
• Report

Mammogram, ultrasound
Ultrasound Mammogram CT etc.
Age, nodule, Family history
23
Triple assessment of breast cancer
Hurt C, Patkar V, Steele R, Rose T, Fox J (in
preparation) Steele R, Fox J Proceedings of
European conference on AI in Medicine, 2003
24
From theory to design
Specialist services
General hospitals
PROforma specification language
Research/ centres of excellence
Primary care
Home and self care
25
From process model to process design PROforma
Decisions
Plans
J Fox and S Das, AAAI and MIT Press 2000
26
A triple assessment model
27
Formal design of clinical processes
28
Tallis NTstandards and interoperability
Care pathways (e.g. PROforma, ASBRU, GLIF)
29
Some PROforma applications

• ACL
• CAPSULE, major improvement in GP prescribing (BMJ
  1997)
• CADMIUM, helping radiographers interpret
  mammograms (Br J Radiol, 1998)
• This Quinquennium
• RAGs genetic risk assessment improved capture of
  family history by GPs and high level of
  acceptability (BMJ 2000)
• CREDO, patient journey for breast cancer
• Retrogram (developed by InferMed for Roche).
• HIV-1 genotyping interpreted by software greatly
  improved the virological outcome when added to
  the clinical information for decisions on
  changing antiretroviral therapy. AIDS 2002
• Promising model for post-genomic medicine

30
InferMed Projects 2003-4 (Arezzo)

• British Medical Journal - Evidence without
  Effort project with Clinical Evidence.
• Brown University Pain Management, Long term
  care of the elderly.
• Mater Misericordiae Dublin 2 electronic
  guidelines
• New Zealand Ministry of Health Diabetes
  Management in General Practice
• Eclipsys Inc. - Remarketing Agreement for AREZZO
  in the US. Also supporting PROforma within HL7
  organisation.
• F Hoffman-La Roche
• Retrogram
• Trans-national study (ORAMA) on Acute Renal
  Anaemia using CDS.
• Pfizer 3 Post-Op pain management guidelines
  (PROSPECT)
• IBA Healthcare Sepsis in ICUs (Pilot project in
  Singapore)
• Accenture advocating adoption of AREZZO within
  NHS

31
Topics

• Motivation improving care
• Modelling and rational design of clinical
  processes
• OpenClinical
• The CREDO trial

• Principles of cognitive systems
• Natural language interfaces
• From standard pathways to individual care plans
• Modelling the cognitive organisation

32
Fostering quality, safety and ethical practice in
clinical knowledge management

• OpenClinical

33
www.OpenClinical.org
34
OpenClinical.netA new way of disseminating
knowledge?
35
R Steele et al, Proc. AI in Medicine Europe, 2003
36
OpenClinical.net Developments

• Initial publications
• Breast cancer demonstrator
• NICE BrCa genetics guideline
• Triple assessment
• Non-cancer demonstrators (with partners)
• Repository construction
• University of Negev (Israel)
• Interoperability techniques
• MEDAL Medical Algorithms group (Houston)
• Funding (EU?)

37
Topics

• Motivation improving care
• Modelling and rational design of clinical
  processes
• OpenClinical
• The CREDO trial

• Principles of cognitive systems
• Natural language interfaces
• From standard pathways to individual care plans
• Modelling the cognitive organisation

38
The CREDO trial
39
Supporting the cancer journey
40
CREDO, phase 1 (2002

• Pathway and decision model developed with
  clinical colleagues (Alyssa Alabassi, Vivek
  Patkar, Matt Williams)

CREDO
41
CREDO, phase 1

• Total of 222 clinical, patient and communication
  services identified about 65 require
  professional decisions
• Software demonstrator constructed
• A number of prototypes implemented in PROforma
  and web technologies
• Pilot trials initiated

42
Example genetic risk assessment with imaging
43
adding value
44
CREDO content

• Complete
• GP referrals (Leicester trial www.infermed.com)
• Triple assessment (pilot)
• Planning care of at-risk women (REACT, pilot)
• NICE guideline genetic risk (to be validated by
  SCHARR)
• Gail tissue density model/mammogram
• Cancer staging
• In development
• Support for multi-disciplinary meeting
• Adjuvant therapy
• Palliative care
• Therapy planning (REACT/Polyphony)

NB The CREDO PROforma model May be a valuable
research resource in itself, both as a possible
model for other cancers and as a corpus for
investigating formal questions in logic and
knowledge representation
45
65 Decisions

• If compliance with best practice is 99 then 50
  of women will get perfect care (0.9965)
• If compliance is 95 then 3 of women will get
  perfect care (0.9565)
• Our results from several studies suggest actual
  departures from quality standards are between 10
  and 30!

46
I cant believe we really do that badly for our
patients (David Miles, Guys)

• Paper patients are not representative of
  clinical reality
• Checks and balances prevent harmful consequences
• Younger, inexperienced staff might benefit but
  this is just a training issue

47
(No Transcript)
48
Phases of the CREDO trial

• Phase 1 activities (2003-4) included
• Service model, technology demonstrator and
  clinical prototypes
• Detailed development of Work-up and
  Treatment-planning
• Pilot trials with paper patients
• Phase 2 activities will include
• 65 decisions study (what is really going on?)
• Multi-centre trials of individual clinical
  services
• Create necessary infrastructure for integrated
  pathway electronic patient records, messaging
  and integration (e.g. medical imaging)
• Phase 3 activities will include
• Controlled trial of integrated services

49
Topics

• Motivation improving care
• Modelling and rational design of clinical
  processes
• OpenClinical
• The CREDO trial

• Principles of cognitive systems
• Natural language interfaces
• From standard pathways to individual care plans
• Modelling the cognitive organisation

50
From information systems to cognitive systems
the main issues

• David Glasspool

51
Key issues

• Flexibility
• Adaptability
• Robustness

52
Cognitive systems

• "Cognitive systems are natural or artificial
  information processing systems, including those
  responsible for perception, learning, reasoning,
  decision-making, communication and action
• UK DTI Technology Foresight
• A cognitive system is one that, among other
  things,
• can reason with appropriately represented
  knowledge
• can learn from its experiences
• can explain itself and can accept direction
• can be aware of its own behavior and reflect on
  its own capabilities
• can respond in a robust manner to surprises.
• DARPA

53
Domino as a cognitive system

• Domino
• Abstraction from clinical applications.
• Outline cognitive system / cognitive theory.
• Correspondences with theories in AI and
  Neuroscience.
• Norman and Shallice theory (neuroscience)
• Processes well-specified, interactions poorly
  specified.
• Existing specification maps cleanly on to Domino.
• Proving to be productive for the original theory
  for view of Domino as cognitive system.

54
Main Processes
Perspective has changed Initially viewed as a
framework for identifying tasks of interest in
clinical processes moved to dynamic view a
theory of processes involved in intelligent
behaviour.
55
1. Goals Issues

• Intelligent behaviour is guided by objectives,
  not rigid pathways.
• Respond flexibly in adverse/changing
  circumstances
• When actions fail, plan a safe alternative.
• Rationale of advice must be explicit.
• Ground explanations in goals.
• Discuss negotiate with humans/other agents.
• Critique clinicians' decisions.
• What formal properties should goals have?
• What formal constraints must be placed on them?

56
1. Goals Approach

• Syntax (based on CREDO analysis)
• ltSituationgt If patient presents with symptoms
  of possible cancer
• ltVerb phrasegt refer patient to
• ltNoun phrasegt oncologist
• ltRestrictionsgt within two weeks
• ltPolicygt obligatory
• ltRationalegt Requirement of NHS cancer
  referral policy
• Semantics (e.g. Winikoff)
• Known (goals must be explicit if an agent is to
  reason about them)
• Consistent (goals that conflict should not be
  adopted)
• Persistent (goals exist so long as their
  success conditions are not satisfied)
• Unachieved (a goal is dropped precisely when its
  success conditions are true)
• Possible (a goal is dropped with failure when
  it becomes impossible to achieve)

57
2. Planning

• Planning is a well established problem in AI
• STRIPS, POP, LCP, HTNs
• Medicine raises challenging issues
• eg uncertainty, change, regulation.
• What types of planning are useful in cognitive
  systems?
• How should plans be represented so that they can
  be
• Created executed?
• Reasoned about?
• Explained?
• Justified?
• Approach convergence of ideas (Neuroscience, AI,
  Agents) investigation of cognitive mechanisms.

58
3. Plan Execution

• One area where the unintelligence of machines
  is very obvious.
• What does it mean to execute a plan
  intelligently?
• Generalisation Not having to specify every last
  detail.
• Flexibility Ability to change details to suit
  circumstances while preserving overall
  intentions.
• Resilience Ability to detect when situation is
  not developing as expected and either adapt the
  plan or request re-planning (or other problem
  solving).
• Approaches
• Flexible plan execution in AI (LCP/HTN, BBAI).
• Goal/action layered hierarchies, and/or trees.
• Cognizant failure planned recovery from failure.
• Action plans in neuroscience/cognitive science.

59
4. Monitoring

• Unmonitored behaviour is dangerous.
• Reflection/monitoring key to adaptability
  resilience.
• Approaches in
• Neuroscience (monitoring processes in
  neurological syndromes).
• AI (e.g. alarm systems in cognitive
  architectures).
• Monitoring processes need to
• Interpret, not just sense, state of world.
• Take account of goals of actions, not just
  effects.
• Take account of potentially conflicting agendas
  of different agents need for negotiation.

60
5. Logic of argument

• Logic and probability theory are classical ways
  to address inference and decision making.
• With emergence of AI, shortcomings become
  evident.
• Argument logic
• Radically new approach.
• Unifies traditional concepts.
• Practical foundation for cognitive systems.
• Ideas from e.g. law, medicine, software
  engineering, education.
• ACL contribution novel approach to decision
  theory.

61
Summary

• Basic science agenda for next quinquennium
• Cognitive systems, emphasising flexibility,
  adaptability robustness through
• Formalising goals (flexibility, robustness)
• Planning on goals (adaptability)
• Decision making (robustness)
• Plan execution (flexibility)
• Monitoring (adaptability robustness)

62
Topics

• Motivation improving care
• Modelling and rational design of clinical
  processes
• OpenClinical
• The CREDO trial

• Principles of cognitive systems
• Natural language interfaces
• From standard pathways to individual care plans
• Modelling the cognitive organisation

63
Natural language interfaces
64
Semantics of medical knowledge
65
Applying the knowledge
Spoken/typed input
66
Applying the knowledge
Speech recogniser (commercial)
Spoken/typed input
Finite state machine for low level response
selection
High level dialogue manager
67
Extending HOMEY

• Acquire new concepts and words (inserting items
  into the ontology)
• Giving clinical orders (generating PROforma
  process specifications)
• Improved explanation capabilities (reflecting on
  agents beliefs and their justifications, goals
  and intentions)

68
Topics

• Motivation improving care
• Modelling and rational design of clinical
  processes
• OpenClinical
• The CREDO trial

• Principles of cognitive systems
• Natural language interfaces
• From standard pathways to individual care plans
• Modelling the cognitive organisation

69
Polyphony From standardised pathways to personal
care plans
70
Individualising care

• A 34 year old woman with a confirmed BRCA1
  mutation, who has had one child.
• Probability (risk) of developing breast cancer
  and some factors directly affecting this are
  reasonably well known for this group
• Many are only qualitatively known or suspected

REACT
71
REACT and Polyphony
72
Polyphony Individualising care
73
The National Programme for ITs concept of the
integrated care pathway
74
Topics

• Motivation improving care
• Modelling and rational design of clinical
  processes
• OpenClinical
• The CREDO trial

• Principles of cognitive systems
• Natural language interfaces
• From standard pathways to individual care plans
• Modelling the cognitive organisation

75
Modelling the cognitive organisation
76
Organisations as networks of cognitive agents
COGENT demonstration
77
(No Transcript)
78
Managing referrals an experiment in silico

• Baseline patients seen in the order that they
  are referred to each centre. Patient joins the
  back of the waiting list at each centre,
  resulting in additive delays.
  
  
  
• Provisional booking all downstream options are
  given a provisional booking after the referral
  decision is taken this is cancelled and returned
  to pool if not needed.
• Provisional booking with urgency assessment
  provisional appointments made as before. If
  assessment agent decides to refer on it assesses
  urgency. If non-urgent patient given her
  provisional slot. If urgent and an earlier slot
  she is offered that.

79
In silico referral experiments data
80

State of the art review and require-ments
Formal semantics
ASPIC Novel agents and other technologies based
on argumentation logics EU funded CRUK
(scientific coordination) UPC Barcelona (SP) CUNY
(USA) U Liverpool (UK) U Utrecht (NL) U Surrey
(UK) IRIT Toulouse (FR) NAVUS (DE) ZEUS/Logic DIS
(GR)
New concepts and experiments
Software engineering
Application demonstrators (e.g. medical)
Dissemination, Exploitation
Project management
81
Focus successor to PROforma

• T Transactions and messages
• A Autonomy with Accountability
• R Roles and Responsibilities
• G Goal based planning and scheduling
• E Events and time
• T extended ontology of Tasks

82
Posters

• ASPIC
• CREDO

• Tallis and OpenClinical
• REACT and Polyphony

83
Topics

• Motivation improving care
• Modelling and rational design of clinical
  processes
• OpenClinical
• The CREDO trial

• Principles of cognitive systems
• Natural language interfaces
• From standard pathways to individual care plans
• Modelling the cognitive organisation

84
Remarks from last review

• Research strategy
• Exploring the nature, applications and
  implications of the domino provides a broad focus
  for ACL work
• The virtuous circle is itself a strategy CRUK
  is both a mission-oriented organisation and one
  that values basic research and we think this
  works well

85
Remarks from last review

• Working with senior clinicians
• Our work has reached a level of maturity where we
  have something to interest clinicians
• The CREDO trial is an important opportunity to
  translate research into patient benefit and has
  the interest of senior oncologists
• Funding the trial is a high priority

86
Supervised entity
GENISYS (NAVUS, Ravensburg)
Delegate agent
Core
GENESYS middleware
Supervisor agent
operator
Prototype
Platform
Concept
87
Supervised entity
GENISYS Tallis TalliSYS
Delegate agent
Core
GENESYS middleware
operator
Prototype
Platform
Concept
88
Summary of short to medium term objectives
89
Applications projects

• CREDO,
• Priority is to secure funding for trial
• Continue development of clinical services
• Develop integrated system and initiate
  multi-centre trial
• OpenClinical.net
• Develop repository and carry out trial
• Secure funding for collaborative development

90
Informatics projects

• Tallis NT emphasis on standards,
  interoperability and integration.
• REACT and Polyphony improved planning and
  scheduling
• HOMEY spoken dialogue more natural user
  interfaces exploiting ontological knowledge and
  task constraints

91
Basic research

• Continuing investigation of the main functions of
  the domino model as a basis for understanding
  natural and artificial cognitive systems
• ASPIC Non-classical logics as a general
  foundation for novel agents
• TARGET successor to PROforma

92
Assessment of approach

• The ACLs theoretical and interdisciplinary
  approach spreads resources thinly but is highly
  productive
• The continuity provided by CRUK core funding is
  very helpful, lets us take a long view, though
  some refinements would be valuable

93
(No Transcript)
94
Project Tallis NTstandards and interoperability
Care pathways (e.g. PROforma, ASBRU, GLIF)
95
BusinessProcessModelling(BPMN)
96
Mapping BPMN toPROforma
T Benson and R PageNPfIT, 2004
97
Translating research into the clinic
Lab data, Ontologies etc.

• A significant
• opportunity?

Research trials
Standard care pathways
Individual patient profiles
Individual care plans
Patient services
Clinical services
Administrative services
98
opportunities for collaboration (but
unfortunately not funding)

• SCHARR
• will review PROforma version of NICE genetics
  guideline, possible larger collaboration on
  further cancer guidelines
• National electronic Library for Health (genetics)
• Proposed joint development based on RAGs
• National Patent Safety Agency
• Creation and maintenance of safety case methods
  and technologies (RED box)
• NCRI informatics task force
• Translation project in discussion
• SPSS
• Agreed to support development of open data portal