SweetP - answers and questions from the front line Dr Ian Wells Royal Surrey County Hospital Guildford Presented at the Integrated Care Records: Problems and Solutions Workshop Edinburgh, 11th to 12th December 2003 1 Answers and questions • Clinical Computing at the Royal Surrey (RSCH) • SweetP - principles and practice • Answers - and future plans • Questions - are we missing the point? • from how to why and what ... and please savage gently!! 2 Clinical Computing at the Royal Surrey County Hospital 3 Computing in a hospital? • section in Medical Physics Department • scientific - supporting work of department • network,Varis, interfacing, image server • clinical - most ‘customers’ outside department • not related to IM&T Department • employ clinical scientists or equivalent 4 Clinical Computing • development - clinical databases (since 1987) • research - IKBS and medical reasoning • teaching - UniS Dept of Computing MSc • strategy - ASSIST PDC, CCL SAG 5 Clinical Databases Operational Development / future Clinical Results Server (CRS) West Surrey Hepatitis C Service Diabetes and Endocrinology (SweetP) Orthodontics Cancer Network (CanCAS) Endoscopy Nuclear Medicine (NMR) Audiology Urology Seeds (SPOC) Patient information leaflets RSCH Intranet Server Macmillan Cancer Nurses Clinical Investigation Unit Medical Photography and Graphics Trace Element Laboratory (MacPath) Pathology Results Archive 6 EPR strategy • fill gaps OASIS cannot reach • tightly integrated with OASIS (4D - Oracle) • inhibit spread of standalone Access databases • satisfy needs of IT enthused consultants • OASIS remains master system 7 Development strategy • 4D software tools [1] since 1987 • 4D Co-operative since 2002 • all software open source • modular and moving to components • clinical database shell for new systems 8 Clinical database shell Specialized software components which only apply to a single clinical application (e.g. importing fundus images) Clinical data layer - specific to actual clinical application (SweetP, HepC, Lipids ...) Generic data layer - patient specific information and other data items which are common to all clinical applications Generic software components including attribute/value manager, PAS and pathology interfaces, statistics, advanced reporting, inference engine and web access Foundation software layer for common presentation and operation, help system, password control, basic reporting 4th Dimension 2003 software for development and deployment Operating System (UNIX, Mac OS, Windows) 9 SweetP - principles & practice 10 SweetP background • management of diabetic clinic • developed for Prof David Russell-Jones • single use since 1999, all doctors since 2001 • over 4900 clinic attendances recorded • new version using shell to support NSF 11 Failure criteria • synchronize patient details with OASIS • generate clinic letter without dictation • provide individual and overall statistics • easy to use for doctors and secretaries 12 SweetP in the diabetic clinic even the Professor can use it ... 13 SweetP in the office and the secretary is smiling again! 14 Comparative statistics Mean BMI BP syst BP diast HBA1c Chol Trigs All 28.7 142.3 79.2 8.4 5.1 2.1 Male 28.5 142.3 79.5 8.3 4.9 2.2 Female 28.9 142.4 78.9 8.6 5.2 2.0 First 6 months 28.3 142.7 79.4 8.5 5.0 2.4 Last 6 months 28.9 141.4 78.4 8.2 4.9 2.0 15 HBA1c distribution HBA1c 700 600 500 400 300 200 100 0 4 6 8 10 12 14 16 18 whole clinic - but could be for single doctor for peer review 16 ‘Front cover of notes’ 17 Routine clinic visit 18 Answers - and future plans 19 Why did SweetP succeed? • met failure criteria • robust and easy to use • in-house development leads to local ownership • research initiative so free to users • prepared to discard prototypes 20 SweetP prototypes • data set derived from British Diabetic Association recommendations • data set reflecting existing ‘paper database’ • data set based on consensus of local diabetes consultants at Health Authority meeting • which one succeeded - what does this tell us? 21 Well behaved data? • selection lists for coded items • mapping to multiple external codes • field labels and colours • multiple ranges for numeric data • rule-based actions where needed 22 EAV module • Entity - Attribute - Value manager • inspired by MYCIN [2] • user control of attributes and values • entity definition stays with developer! • integral component of clinical systems shell 23 24 25 26 Questions are we missing the point? 27 Doctors talk to patients ... 28 Pass around information ... 29 ... and make decisions! 30 Better information - better medicine? “Medicine is the art of making acceptable decisions in an imperfectly understood problem space using insufficient and often erroneous information” With acknowledgment to Edward Shortliffe 31 How do doctors reason? • Harvard studies by Elstein in 1978 [3] • 24 doctors and 3 specially selected cases • grouped into ‘experts’ and ‘remainder’ • retrospective analysis of management of cases • focus on acquisition of cues + reasons for errors 32 Backwards not forwards! • hypothetico-deductive approach to diagnosis • pure inductive reasoning not used • up to five hypotheses in memory • Miller’s ‘magical number’ 7 +/- 2 (1956) • 3-point weighting of uncertainty • 7-point experiment showed no better results • experts better at data gathering & interpretation 33 Use computers with caution • doctors are human, tend to reason backwards and are limited by Miller’s number • the order and emphasis of data items is important in both entry and review • departure from established practice could lead to early discarding or merging of the correct hypothesis and thus the wrong decision 34 Food for thought • established ‘paper database’ may represent best practice in clinical information management • we should not rely solely on the doctor’s intuition for our understanding of the medical decision-making processes • are our suppliers aware of the question - let alone the answers? • we ignore the cognitive aspects of computing at our peril (perception and cognition) 35 Neisser’s cyclic model of perception Top down Expected features Deductive Perceptual model Environment Bottom up Feature analysis Inductive 36 Conclusion “The (medical) deductive process is not automatic, even though experienced practitioners can make it seem to be” Elstein (Harvard project [3] page 111) 37 Acknowledgments • Richard Harper - getting the RSCH involved • Alex Taylor - fieldwork and encouragement • Jose Spring - keeping us in order • Rob Scott et al - invitation to participate 38 References 1. 4D software tools: www.4duk.com 2. Buchanan BG, Shortliffe EH eds, ‘Rule-Based Expert Systems: the MYCIN Experiments of the Stanford Heuristic Programming Project’, Addison-Wesley 1985. 3. Elstein AS, Shulman LS, Sprafka SA, ‘Medical Problem Solving: an Analysis of Clinical Reasoning’, Harvard University Press, 1978 39 Contact details Dr Ian Wells Clinical Computing Section Department of Medical Physics Royal Surrey County Hospital Guildford Surrey GU2 7XX UK Tel: +44 (0)1483 464 039 Email: ian.wells@royalsurrey.nhs.uk Also: Department of Computing, University of Surrey 40 The end 41