1. Applying AI temporal reasoning techniques to Clinical Guidelines Luca Anselma%, Paolo Terenziani*, Stefania Montani*, Alessio Bottrighi* %DI, Università di Torino, Corso Svizzera 185, 10149 Torino, Italy Phone: +39 011 6706821 E-mail: anselma@di.unito.it *DI, Univ. del Piemonte Orientale “Amedeo Avogadro” Spalto Marengo 33, 15100 Alessandria, Italy Phone: +39 0131 360174 E-mail: {terenz,stefania,alessio.bottrighi}@mfn.unipmn.it - Introduction - Temporal Constraints in Clinical Guidelines: new issues - An extension to AI Temporal Reasoning Techniques - Conclusions

2. Introduction Many different computer systems managing clinical guidelines (e.g., Asgaard, GEM, Gliff, Guide, PROforma,…) Different roles: -support -critique -evaluation -education -…... Clinical guidelines are a means for specifying the “best” clinical procedures and for standardizing them Adopting (computer-based) clinical guidelines is advantageous

3. GLARE (GuideLine Acquisition Representation and Execution) -Joint project: Dept. Comp. Sci., Univ. Alessandria (It): P. Terenziani, S.Montani, A.Bottrighi Dept. Comp. Sci., Univ. Torino (It): L.Anselma,G.Correndo Az. Osp. S. Giovanni Battista, Torino (It): G.Molino, M.Torchio -Domain independent (e.g., bladder cancer, reflux esophagitis, heart failure) -User-friendly (limited number of primitives)

4. GLARE Representation Formalism

5. Temporal Constraints in Clinical Guidelines Temporal constraints are an intrinsic part of clinical knowledge (e.g., ordering of the therapeutic actions) (1) “standard” AI constraints -duration of actions (min / max) -qualitative constraints (e.g., before, during) -delays (min / max) Explicit (duration of actions, delays between actions) or induced from the control relations (ordering of actions)

6. Temporal Constraints in Clinical Guidelines (2) Implicit constraints induced from the part-of relations -standard “containment” constraints (!?)

7. Temporal Constraints in Clinical Guidelines (3) Explicit constraints on repeated actions (Ex. 2) Intrathecal methotrexate must be administered 7 times during 88 weeks, never less than 10 weeks apart or more then 14 weeks apart. -NOTICE: the number of repetitions may be unknown (Ex. 3) Give acetaminophen twice a day until the fever has gone. -NOTICE: complex interplay with part-of (nested repetitions) (Ex. 1) The therapy for multiple mieloma is made by six cycles of 5-day treatment, each one followed by a delay of 23 days (for a total time of 24 weeks). Within each cycle of 5 days, 2 inner cycles can be distinguished: the melphalan treatment, to be provided twice a day, for each of the 5 days, and the prednisone treatment, to be provided once a day, for each of the 5 days. These two treatments must be performed in parallel.

8. Temporal Constraints in Clinical Guidelines (4) Distinction between -constraints between “classes” of actions -constraints between “instances” of actions  Inheritance of constraints  Predictive role of “classes”

9. Temporal Constraints in Clinical Guidelines OUR CHALLENGE: REPRESENTING AND REASONING WITH ALL SUCH CONSTRAINTS IN AN INTEGRATED WAY NO OTHER APPROACH IN THE LITERATURE CHALLENGING EXTENSION TO “STANDARD” APPROACHES

10. Managing Temporal Constraints: the Problem DESIDERATA for Temporal Reasoning Algorithms - tractability  “reasonable” response time - correctness  no wrong inferences - completeness  reliable answers DESIDERATA for the Representation formalism - expressiveness  capture most temporal constraints in GL TRADE-OFF!

11. Our solution: two layered approach (1)HIGH-LEVEL Language -part-of -Instance of -Repetition Repetition(A, R 1, R 2, …, R n ), R i =, repConstraints i  fromStart(min, max), toEnd(min, max), inBetweenAll(min, max) inBetween((min 1, max 1 ), …, (min nRep-1, max nRep-1 )), Conditions i  onlyIf(B), while(B)

12. Our solution: two layered approach (1)LOW-LEVEL language + Temporal Reasoning Based on STP [Decther et al., 91] Extensions NEEDED!

13. Starting Point: STP framework Conjunctions of b.o.d. constraints c  X-Y  d Floyd-Warshall all-to-all shortest paths algoritm is correct and complete for the STP framework, and operates in O(N 3 ) (where N is the number of variables - time points) It produces the minimal network of the constraints (i.e., the shortest path between each pair of nodes) Can be used to represent distances between points (starting/ending points of actions)

14. Labeled tree of STPs (STPs-tree). STP-tree for the multiple mieloma chemotherapy guideline in Ex. 1. Thin lines and arcs between nodes in a STP represent bound on differences constraints. Arcs from a pair of nodes to a child STP represent repetitions. Sch, Ech, Smc, Emc, Spc, Epc, Sm, Em, Sp and Ep stand for the starting (S) and ending (E) points of chemotherapy, melphalan cycle, prednisone cycle, melphalan treatment and prednisone treatment, respectively. Sm Sp,, [0d,1d] [0d,0d ] [168d,168d] [0d,0d] [5d,5d] [0d,1d] N3 N4 N2 N1 Sch Emc Smc Epc Spc Ep Em Ech, > [5d,5d] [0d,0d] [5d,5d]

15. Consistency checking on STPs-trees ALGO1: temporal consistency of guidelines Top-down visit of the nodes in the STPs-tree For each node in the STPs-tree: function STP_tree_consistency(X: STPNode, RepSpec)) : STP (1)check that the repetition/periodicity constraint is well-formed (i.e., that repetitions nest properly) (2)compute Max, i.e. the maximum duration of a single repetition of X according to RepSpec (3)impose in X that the maximum distance between each pair of points is less or equals Max (4)X  FloydWarshall(X) (5)if X = INCONSISTENT then return INCONSISTENT else return X

16. Consistency checking on STPs-trees ALGO1: Properties Complexity. Considering that the number of nesting levels, in the worst case, is less than the number of classes, the algorithm is dominated by step 4, that is O(C 3 ), where C is the number of actions in the guideline. Property 1. The top-down visit of the STP-tree is complete as regards consistency checking of the constraints in the STP- tree.

17. Temporal reasoning algorithms on STPs-trees classes+instances function integratedConsistency(T : STP-tree, E : executionSTP, NOW) : STP a. (1) check that in the executionSTP there are all and only the instances that the STP-tree predicts to be. (2) Possible missing instances are hypothesized because they may happen in the future; // this step deals with the predictive role of the temporal constraints about classes b. (3) inherit the repetition/periodicity constraints and the temporal (non-periodic) constraints from the classes to the instances; //this step “copies” all temporal constraints to the executionSTP; c. (4) propagate the temporal constraints on the executionSTP, thus obtaining the minimal network; (5) check whether the hypothesized instances expected in the future may actually start in the future (i.e., after NOW) ALGO 2

18. Temporal reasoning algorithms on STPs-trees Properties Complexity. Let us denote with C the number of classes in the STP-tree and with I the number of instances in the executionSTP. The complexity of integratedConsistency procedure is O(max{C 3, I 3 }). Property 2. The integratedConsistency procedure is correct and complete as regards consistency checking of the constraints in the executionSTP and in the STP-tree.

19. Temporal Reasoning for Clinical Guidelines During acquisition, to check consistency (Algo 1) During execution (1) for scheduling the next action (2) for quality evaluation: check whether “classes” constraints have been respected by instances (Algo 2) (3) to support decision making: queries (4) to support decision making: comparing guidelines paths

20. Temporal Reasoning for Clinical Guidelines e.g., for task 4: 1.For each path P i to be compared 2.Hypothesize the existence of an instance of each action in P i which has not been executed yet 3.Apply the algorithm in Fig. 3 to the (executed and hypothesized) actions in P i, to determine the minimal network MN i 4.Retrieve the minimal and maximal duration of P i from MN i

21. Conclusions Temporal constraints are an intrinsic part of clinical guidelines Implicit + explicit constraints NEW CHALLENGING PROBLEMS OUR SOLUTION: -analysis of the trade-off between expressiveness and complexity -definition of STPs-trees -tractable, correct and complete constraint propagation algorithms -additional algorithms, to exploit constraint propagation in the guideline context CONCLUSION: TR can give relevant contributions to the MI field!