Presentation at AIME 2011

341 views

Published on

From time-annotated Clinical Guidelines to Temporal Hierarchical Task Netowrks

Published in: Business, Economy & Finance
  • Be the first to comment

  • Be the first to like this

Presentation at AIME 2011

  1. 1. Careflow Planning: From Time-annotated Clinical Guidelines to Temporal Hierarchical Task Networks Arturo González-Ferrer Annette ten Teije Juan Fdez-Olivares Krystyna Milian 1miércoles 6 de julio de 2011
  2. 2. MOTIVATION • Real requirements by doctors of 7 hospitals in our region (Andalucía) in project Oncotheraper: • Virgen del Rocío (Sevilla) • Virgen Macarena (Sevilla) • Carlos Haya (Málaga) • Torre Cárdenas (Almería) • Reina Sofía (Córdoba) • Complejo Hospitalario (Jaén) • Virgen de las Nieves (Granada) 2miércoles 6 de julio de 2011
  3. 3. MOTIVATION 3miércoles 6 de julio de 2011
  4. 4. MOTIVATION • Doctors in the project need IT support for planning, visualization and execution of a patient long-term treatment in pediatrics oncology 3miércoles 6 de julio de 2011
  5. 5. MOTIVATION • Doctors in the project need IT support for planning, visualization and execution of a patient long-term treatment in pediatrics oncology • They spend much time preparing the patient plan at hand 3miércoles 6 de julio de 2011
  6. 6. MOTIVATION • Doctors in the project need IT support for planning, visualization and execution of a patient long-term treatment in pediatrics oncology • They spend much time preparing the patient plan at hand • These protocols are very constrained by multiple temporal annotations (chemotherapy cycles, delayed synchronizations, ...). 3miércoles 6 de julio de 2011
  7. 7. MOTIVATION • Doctors in the project need IT support for planning, visualization and execution of a patient long-term treatment in pediatrics oncology • They spend much time preparing the patient plan at hand • These protocols are very constrained by multiple temporal annotations (chemotherapy cycles, delayed synchronizations, ...). • We did previous work where we used the BPMN graphical notation to model organizational processes and obtain process instances adapted to specific conditions 3miércoles 6 de julio de 2011
  8. 8. MOTIVATION • Doctors in the project need IT support for planning, visualization and execution of a patient long-term treatment in pediatrics oncology • They spend much time preparing the patient plan at hand • These protocols are very constrained by multiple temporal annotations (chemotherapy cycles, delayed synchronizations, ...). • We did previous work where we used the BPMN graphical notation to model organizational processes and obtain process instances adapted to specific conditions • we found that the capability of BPMN to represent complex temporal constraints is still confusing and incomplete 3miércoles 6 de julio de 2011
  9. 9. MOTIVATION • Doctors in the project need IT support for planning, visualization and execution of a patient long-term treatment in pediatrics oncology • They spend much time preparing the patient plan at hand • These protocols are very constrained by multiple temporal annotations (chemotherapy cycles, delayed synchronizations, ...). • We did previous work where we used the BPMN graphical notation to model organizational processes and obtain process instances adapted to specific conditions • we found that the capability of BPMN to represent complex temporal constraints is still confusing and incomplete • time-BPMN extension (Gagné & Trudel) is only theoretical, not practical 3miércoles 6 de julio de 2011
  10. 10. MOTIVATION • Doctors in the project need IT support for planning, visualization and execution of a patient long-term treatment in pediatrics oncology • They spend much time preparing the patient plan at hand • These protocols are very constrained by multiple temporal annotations (chemotherapy cycles, delayed synchronizations, ...). • We did previous work where we used the BPMN graphical notation to model organizational processes and obtain process instances adapted to specific conditions • we found that the capability of BPMN to represent complex temporal constraints is still confusing and incomplete • time-BPMN extension (Gagné & Trudel) is only theoretical, not practical • We decided to move the modeling stage to using Computer-interpretable Guidelines 3miércoles 6 de julio de 2011
  11. 11. “Hardly any of the existing Clinical Decision Support Systems (CDSS) appear to be aimed at supporting extended clinical workflows, management of information and decision-making in plans that unfold over time” J. Fox, D. Glasspool, V. Patkar, M. Austin, L. Black, M. South, D. Robertson, and C. Vincent. Delivering clinical decision support services: There is nothing as practical as a good theory. Journal of Biomedical Informatics, 43(5):831-843, 2010 4miércoles 6 de julio de 2011
  12. 12. MOTIVATION • We need to represent and reason about actions and decisions, different patient profiles, temporal patterns and scheduling of resources, in order to automatically generate personalized care plans • We need these plans to be visually atractive, and useful for the doctors • We want to deploy these plans for the patient treatment into a workflow engine (i.e. careflow) • This way, both the patients and doctors could follow the treatment using a more efficient, safe, and high-quality healthcare process 5miércoles 6 de julio de 2011
  13. 13. CLINICAL GUIDELINES • Clinical Practice Guidelines: Recommendations on the appropriate treatment and care of people with specific diseases and conditions, on the basis of the best available evidence. • Computer-interpretable Guidelines languages 6miércoles 6 de julio de 2011
  14. 14. CARE PATHWAYS Care Pathways: Aim to model a timed process of patient-focused care, by specifying key events, clinical exams and assessments to produce the best prescribed outcomes, within the limits of the resources available, for an appropriate episode of care • It aims to reduce the patient’s stay time in the hospital, delivering a more efficient care process 7miércoles 6 de julio de 2011
  15. 15. AIMS AND GOALS • Starting from the recommendations of a Computer Interpretable Guideline for a specific disease, we want to obtain a Care Pathway: Clinical Guidelines 8miércoles 6 de julio de 2011
  16. 16. AIMS AND GOALS • Starting from the recommendations of a Computer Interpretable Guideline for a specific disease, we want to obtain a Care Pathway: Clinical Guidelines ? 8miércoles 6 de julio de 2011
  17. 17. AIMS AND GOALS • Starting from the recommendations of a Computer Interpretable Guideline for a specific disease, we want to obtain a Care Pathway: Care Pathways Clinical Guidelines ? 8miércoles 6 de julio de 2011
  18. 18. AIMS AND GOALS • Starting from the recommendations of a Computer Interpretable Guideline for a specific disease, we want to obtain a Care Pathway: Care Pathways Patient Profile Clinical Guidelines ? 8miércoles 6 de julio de 2011
  19. 19. AIMS AND GOALS • Starting from the recommendations of a Computer Interpretable Guideline for a specific disease, we want to obtain a Care Pathway: Care Pathways Patient Profile Clinical Guidelines Temporal Patterns ? 8miércoles 6 de julio de 2011
  20. 20. AIMS AND GOALS • Starting from the recommendations of a Computer Interpretable Guideline for a specific disease, we want to obtain a Care Pathway: Care Pathways Patient Profile Clinical Guidelines Temporal Patterns ? Resources 8miércoles 6 de julio de 2011
  21. 21. AIMS AND GOALS • Starting from the recommendations of a Computer Interpretable Guideline for a specific disease, we want to obtain a Care Pathway: Care Pathways Patient Profile Clinical Guidelines Temporal Patterns ? Resources Visual Plan 8miércoles 6 de julio de 2011
  22. 22. AIMS AND GOALS • Starting from the recommendations of a Computer Interpretable Guideline for a specific disease, we want to obtain a Care Pathway: Care Pathways Patient Profile Clinical Guidelines Temporal Patterns ? Resources Visual Plan Workflow 8miércoles 6 de julio de 2011
  23. 23. ASBRU • It is an XML-based, time-oriented, CIG language, used to embody CPGʼs and protocols as skeletal plans. • Each one of these skeletal plans consists of a plan-body that can be composed of: • subplans (set of steps in parallel or sequentially), • cyclical-plans (repeated several times), • plan-activations (a call to another plan) or • user-performed steps (a specific action performed by the user). • In addition, time-annotated conditions can be attached for the selection of plans 9miércoles 6 de julio de 2011
  24. 24. HTN PLANNING • HTN planning has shown to be very useful on practical human-centric domains, for the generation of customized plans (tourism, e-learning, emergencies, ...). • HPDL: It is HTN extension of PDDL language (derived from first-order logic), organized as: planning domain planning problem compound tasks hierarchy of objects primitive actions decomposition methods preconditions initial state predicates, functions set of goals 10miércoles 6 de julio de 2011
  25. 25. Computer Interpretable Guidelines are very good for Representation of actions and decisions in Clinical Processes, but they need a counterpart Knowledge Reasoning step to be used as a basis for obtaining Care Pathways 11miércoles 6 de julio de 2011
  26. 26. WHAT TO DO? • Temporal HTN planning is appropriate for the aim of generating care pathways • J. Fdez-Olivares et al. Supporting clinical processes and decisions by hierarchical planning and scheduling Computational Intelligence, 2011 • But : modeling HTN domains directly can be really complex. Sometimes, an Art! • We need a higher level language for representation of the guideline: • Let use Asbru • Then what? • We need some Knowledge Engineering 12miércoles 6 de julio de 2011
  27. 27. WHAT TO DO? • Temporal HTN planning is appropriate for the aim of generating care pathways • J. Fdez-Olivares et al. Supporting clinical processes and decisions by hierarchical planning and scheduling Computational Intelligence, 2011 • But : modeling HTN domains directly can be really complex. Sometimes, an Art! • We need a higher level language for representation of the guideline: • Let use Asbru Asbru to HPDL? • Then what? • We need some Knowledge Engineering 12miércoles 6 de julio de 2011
  28. 28. What is the most important aim of Knowledge Engineering? Turning the process of constructing Knowledge Based Systems from an Art into an Engineering Discipline, using better methodological approaches [Studer et. al, 1998] 13miércoles 6 de julio de 2011
  29. 29. METHODOLOGICAL APPROACH • Experts describe the CPG in text format • Knowledge Engineers model it using a CIG language (e.g. Asbru) 14miércoles 6 de julio de 2011
  30. 30. METHODOLOGICAL APPROACH • Experts describe the CPG in text format • Knowledge Engineers model it using a CIG language (e.g. Asbru) • Automatically transform the Asbru representation into an HTN planning domain model, considering time constraints 14miércoles 6 de julio de 2011
  31. 31. METHODOLOGICAL APPROACH • Experts describe the CPG in text format • Knowledge Engineers model it using a CIG language (e.g. Asbru) • Automatically transform the Asbru representation into an HTN planning domain model, considering time constraints • Interprete the domain to obtain a Care Pathway for a patient, using the available resources 14miércoles 6 de julio de 2011
  32. 32. METHODOLOGICAL APPROACH • Experts describe the CPG in text format • Knowledge Engineers model it using a CIG language (e.g. Asbru) • Automatically transform the Asbru representation into an HTN planning domain model, considering time constraints • Interprete the domain to obtain a Care Pathway for a patient, using the available resources • Deploy the Care Pathway obtained into a BPM engine for ubiquitous execution. 14miércoles 6 de julio de 2011
  33. 33. METHODOLOGICAL APPROACH • Experts describe the CPG in text format • Knowledge Engineers model it using a CIG language (e.g. Asbru) • Automatically transform the Asbru representation into an HTN planning domain model, considering time constraints • Interprete the domain to obtain a Care Pathway for a patient, using the available resources • Deploy the Care Pathway obtained into a BPM engine for ubiquitous execution. 14miércoles 6 de julio de 2011
  34. 34. METHODOLOGICAL APPROACH • Experts describe the CPG in text format • Knowledge Engineers model it using a CIG language (e.g. Asbru) • Automatically transform the Asbru representation into an HTN planning domain model, considering time constraints • Interprete the domain to obtain a Care Pathway for a patient, using the available resources • Deploy the Care Pathway obtained into a BPM engine for ubiquitous execution. already done (see bibliography) 14miércoles 6 de julio de 2011
  35. 35. METHODOLOGICAL APPROACH • Experts describe the CPG in text format • Knowledge Engineers model it using a CIG language (e.g. Asbru) • Automatically transform the Asbru representation into an HTN planning domain model, considering time constraints • Interprete the domain to obtain a Care Pathway for a patient, using the available resources • Deploy the Care Pathway obtained into a BPM engine for ubiquitous execution. already done (see bibliography) work in progress 14miércoles 6 de julio de 2011
  36. 36. MAPPING ASBRU TO HPDL • Why is this useful? • HTN planning is capable to represent clinical protocols and reason to obtain a pathway: • not only representing, but also managing complex temporal patterns • representing custom constraints (e.g. resources) and reasoning about them • adapted to the specific conditions of the patient profile • very useful in low-frequency scenarios, where BPM tools could be interesting • CIG languages more user-friendly for knowledge representation • not only for humans, also for managing them in a computer • some of them use graphical notations, that are underneath stored as XML 15miércoles 6 de julio de 2011
  37. 37. HIGH OR LOW FREQUENCY DOMAIN? ❖ Traditional use of Asbru in high-frequency domains (e.g. intensive care units) Asbru Guideline Time- Asbru annotated Interpreter input data What to do for every ICU input variable? 16miércoles 6 de julio de 2011
  38. 38. HIGH OR LOW FREQUENCY DOMAIN? ❖ We want to use it for low-frequency domains (long/medium-term care plans) Asbru Guideline Translator Patient Conditions Long-term Guideline in IACTIVE Care Hospital HPDL planner Pathway Resources Deliberative Reasoning 17miércoles 6 de julio de 2011
  39. 39. ASBRU & HPDL: SIMILARITIES • Both follow a similar Task Network Model (TNM) • This is also valid for other CIG languages (not only Asbru) • Asbru skeletal plans similar to HTN compound/primitive operators. • Both are able to represent multiple task ordering schemas • Both are able to represent multiple temporal patterns (e.g. Allen’s T.L. ) 18miércoles 6 de julio de 2011
  40. 40. MAPPING ASBRU TO HPDL Asbru HPDL 19miércoles 6 de julio de 2011
  41. 41. MAPPING ASBRU TO HPDL Asbru HPDL plan-activations A B task calls activate B 19miércoles 6 de julio de 2011
  42. 42. MAPPING ASBRU TO HPDL Asbru HPDL plan-activations A B task calls activate B A1 cond 1 A1 subplans parallel conditional compound tasks A A A2 cond 2 A2 if-then-else task methods A A1 A2 sequence 19miércoles 6 de julio de 2011
  43. 43. MAPPING ASBRU TO HPDL Asbru HPDL plan-activations A B task calls activate B A1 cond 1 A1 subplans parallel conditional compound tasks A A A2 cond 2 A2 if-then-else task methods A A1 A2 sequence user-performed A durative-action 19miércoles 6 de julio de 2011
  44. 44. MAPPING ASBRU TO HPDL Asbru HPDL 20miércoles 6 de julio de 2011
  45. 45. MAPPING ASBRU TO HPDL Asbru HPDL duration A ?start, ?end, ?duration time annotations start end <,>,<=,>= 20miércoles 6 de julio de 2011
  46. 46. MAPPING ASBRU TO HPDL Asbru HPDL duration A ?start, ?end, ?duration time annotations start end <,>,<=,>= time-annotated references to plans 28 days after (ref. points as plan A B temporal landmarks pointers) (all allen’s relations) 20miércoles 6 de julio de 2011
  47. 47. MAPPING ASBRU TO HPDL Asbru HPDL duration A ?start, ?end, ?duration time annotations start end <,>,<=,>= time-annotated references to plans 28 days after (ref. points as plan A B temporal landmarks pointers) (all allen’s relations) repeat 5 times cyclical plans recursive cyclical task (based on temporal formalism A frequency of Anselma et. al) offset 20miércoles 6 de julio de 2011
  48. 48. RESULTS • Proof of concept: We focused in the protocol for the Hodgkin’s disease. • around 70 text pages • We modeled it with a subset of the Asbru language (using DELT/A) 21miércoles 6 de julio de 2011
  49. 49. RESULTS • We have developed a KE tool for the translation (Asbru2HPDL) • We identified and translated different patterns, not only the hierarchy of tasks: 22miércoles 6 de julio de 2011
  50. 50. RESULTS • The translation only cover the workflow of the care process, but not the modeling of patient profiles and hospital resources Patient Profile ;;definition of predicates in the domain (sex ?patient ?s - gender) (group ?patient ?g - group) ;; instances for patient Alice in the problem (sex Alice M) (group Alice Group3) ;;start date for treatment (startdate Alice "07/11/2011 08:00:00") Resource Constraints (between "07/11/2011 00:00:00" and "07/12/2011 00:00:00" (available John)) (between "07/12/2011 00:00:00" and "07/01/2011 00:00:00" (available Paul)) (between "07/01/2012 00:00:00" and "07/02/2012 00:00:00" (available John)) (between "07/02/2012 00:00:00" and "07/03/2012 00:00:00" (available Paul)) (between "07/03/2012 00:00:00" and "07/04/2012 00:00:00" (available John)) 23miércoles 6 de julio de 2011
  51. 51. RESULTS • A fragment of the generated pathway: • This pathway can afterwards be translated into a representation understandable by a BPM engine • González-Ferrer, A., Fdez-Olivares, J., Sánchez-Garzón, I., & Castillo, L. (2010). Smart Process Management: Automated Generation of Adaptive Cases based on Intelligent Planning Technologies. 8th BPM Conference, Demo Track 24miércoles 6 de julio de 2011
  52. 52. RESULTS • Personalized access for each doctor to his/her list of time-annotated tasks • Patient’s follow-up procedure is supported by a BPM runtime console 25miércoles 6 de julio de 2011
  53. 53. Conclusions • We presented an AI-based methodology to model and operationalize care pathways for oncology treatments • The patterns commonly found in a CIG representation (concretely Asbru) can be translated into a corresponding HTN representation. • The methodology could be used for other CIG languages as well • This is interesting because • Adding patients and resources information, we can obtain a customized care pathway that unfolds over time • CIGs are more user-friendly for the modeling than doing it directly with cumbersome HTN planning languages • We can finally deploy the care plan into a BPM execution engine 26miércoles 6 de julio de 2011
  54. 54. FUTURE WORK • Use and evaluation in a real hospital scenario • Apply to other protocols, which patterns could be also interesting? • Flexibility: we are progressing to include monitoring, plan repair and replanning • Real integration with real HIS and EHR 27miércoles 6 de julio de 2011
  55. 55. see more at ... • presentation @ KR4HC workshop • Dr. Juan Fernández-Olivares • Task Network based modeling, dynamic generation and adaptive execution of patient-tailored treatment plans based on Smart Process Management technologies 28miércoles 6 de julio de 2011

×