A common ground theory of medical decision-making 1: The CREDO stack John Fox Department of Engineering Science University of Oxford and OpenClinical Thanks to … Clinicians Psychologists, Informatics/CS/AI • • • • • • • • • • • • • • • • • • • • • • • • • • Alyssa Alabassi John Bury Robert Dunlop John Emery Marc Gutenstein Andrzej Glowinski Mike O’Neil Vicky Monaghan Vivek Patkar Jean-Louis Renaud-Salis Robert Walton Matt Williams Guy Wood-Gush Andrew Coulson Ioannis Chronakis Subrata Das David Glasspool Omar Khan Paul Krause Simon Parsons Mor Peleg Ali Rahmanzadeh Matt South Rory Steele Paul Taylor Richard Thomson Summary • Medicine is a rich and challenging domain for decision science and decision engineering • It raises major challenges and curiously neglected questions at many levels – theory, technology, applications and more … • The common ground theory aims to provide a general framework in which to – Promote discussion across disciplines – Clarify research questions and – Develop practical solutions • The CREDO stack is a particular instance, but there are many others The borders of the common ground • “Prescriptive” (axiomatic, rational) theories – Lindley “there is only one correct way to take a decision” – EUT, Multicriteria DT, game theory, … and many ad hoc variants • “Descriptive” (empirical, explanatory) theories – Cognitive (Nobel Laureates - Herbert Simon, Daniel Kahneman) – Neuroscience (neuroanatomy, neuropsychology, “hot cognition”) – Ecological (e.g. Gary Klein “naturalistic” theories) • “Practical” (engineering, design) theories – Decisions are often framed and made with respect to standard practice – Decision systems may need to engage with accepted practice Medical motivation: Quality and safety of patient care • UK National health service – Vincent data on medical error in Acute Hospitals • >10% acute hospital admissions in NHS lead to avoidable medical error • US Institute of Medicine – IOM: “To err is human”; “Crossing the quality chasm” – McGlynn: Quality of Health Care Delivered to Adults in the USA Quality of Health in the USA McGlynn NEJM 2003 The CREDO stack Diversity of medical decisions 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. Screening for and classification of hazards; Risk stratification and management; Selection of tests and investigations; Diagnosing the cause(s) of clinical complaints; Selecting treatments and other interventions; Prescribing drugs (routes, dosages, polypharmacy etc.); Referring patient to a colleague Deciding whether a decision is needed; Initiating, adjusting and stopping treatments; Deciding whether earlier decisions are correct or not; if not why not; adjust; reverse, reframe, retake; Diversity of medical decisions 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. Screening for and classification of hazards; Risk stratification and management; Selection of tests and investigations; Diagnosing the cause(s) of clinical complaints; Selecting treatments and other interventions; Prescribing drugs (routes, dosages, polypharmacy etc.); Referring patient to a colleague Deciding whether a decision is needed; Initiating, adjusting and stopping treatments; Deciding whether earlier decisions are correct or not; if not why not; adjust; reverse, reframe, retake; MDM is reason based Refer to specialist colleague if … 1. 2. 3. 4. 5. 6. 7. There is a possible life threatening condition I don’t know what to do or lack sufficient knowledge The NICE clinical guideline says I should Patient is eligible for a research trial Difficult patient, and I can’t resolve issue by myself Patient has asked to be referred Colleague or mentor has suggested I should … MDM is dynamic • Decision-makers must deal with changing and often unpredictable circumstances • Decisions are not just choices, they are points in an evolving narrative (patient and professional) • Common ground theory should address the whole cycle of decision-making: – – – – When is a decision needed? what is the goal of the decision? What knowledge and strategies are relevant? When is it appropriate and safe to commit? When is it necessary to revisit and revise commitments as the situation evolves? Example: cancer care Example: cancer care Example: decisions in context The CREDO stack A common ground theory Goals Beliefs Actions Options Commitments Plans From decision science to decision engineering: the CREDO stack ResearchGate 2014 Example: risk assessment Assess risk Goals Worried, well Beliefs Moderate risk Actions Genetic, statistical & other lines of reasoning Options Population or moderate or high risk Commitments Plans Example: test selection Investigate for possible cancer tr Pain, nodule Goals Beliefs Order Mammogram & ultrasound Actions Mammogram, ultrasound Options Ultrasound Mammogram CT etc. Age, symptoms, … Family history Commitments Plans Reasons and decisions Argument construction Knowledge U Data LA (Claim, Reason, Qualifier) {(Claim, Reason, Qualifier)} Agg (Claim, Modality) Argument aggregation Fox et al ECAI, 1992; UAI 1994; Fox and Das, 2000 Krause et al Computational Intelligence 1995 Uncertainty and arguments • Quantitative [0,1] [-1,+1] {1,2,3,…n} degree of belief (e.g. probability, possibility) bipolar measures (e.g. belief functions) ad hoc weighting of arguments • Qualitative + {+,-} {++,--, +, -} “supporting” arguments “supporting” and “opposing” arguments … plus “confirming” and “excluding” • Modal Linguistic (perhaps, possible, probable, plausible …) Formalising the common ground theory The CREDO stack The knowledge ladder Agents Expert systems, Personal care agents Plans Care pathways, workflows Decisions Rules Descriptions Concepts Symbols Reasons (arguments, evidence, preferences) Alerts, reminders, interpretations Medical facts, Clinical notes Class hierarchies, semantic networks Diseases, Symptoms, Findings, Drugs Terminologies, coding systems The CREDO stack PROforma: Reification into “tasks” Enquiries Goals Actions Beliefs Plan Decision Candidates Actions Commitments Plans Fox et al, MIE 1996; Fox and Das, AI in hazardous applications, MIT Press, 2000 Decision engineering The CREDO stack Applications Care pathways in cardiology UPMC (USA), NHS (NZ) , NHS UK Diagnosis and treatment in endocrine conditions (thyroid, diabetes) AACE (USA) Decision support for general practitioners BPAC (NZ) Triage for common conditions NHS Choices (UK) Supporting the breast MDT- Royal Free Hospital BASO 2008, ASCO 2009, BMJ Open, 2012 Triple assessment of suspected breast cancer Brit J Cancer 2006 Chemotherapy for children with acute lymphoblastic leukaemia Brit J Haematology 2005 Planning care for women at risk of breast/ovarian cancer Methods of Information in Medicine 2004 GP referrals for common cancers MEDINFO 2003 Genotype of HIV+ patients interpretation and selection of antiretrovirals (InferMed, Hoffman la Roche) AIDS 2002 Genetic risk assessment BMJ 1999, 2000 Support for mammographic screening Medical Imaging 1999 Prescribing in general practice BMJ 1997 The CREDO stack Decision support: human interaction Summary • Medicine is a challenging domain for – Understanding human error and expertise – Developing decision theory, empirical science and engineering methods • It raises many important questions and some strangely neglected ones – This will require contributions from many disciplines but there is a high level of fragmentation in decision science • The “domino” is a first draft of a common ground theory, to promote interdisciplinary discussion • The CREDO stack validates the theory to a first approximation demonstrates its practical value