1Monitoring6.872/HST950Peter SzolovitsOutline• Problem of information overload• How to reason with/about time• Interpreting temporal dataProblem• ICU alarms sound roughly every 30seconds, in a typical (full) ICU• Nurse takes ~minutes to resolve alarm• How to resolve?– Ignore (turn off) alarms– Prioritize– Automate– Make alarming algorithms more intelligentTime is Critical• Some systems have no explicit representation oftime– E.g., Internist• <ABDOMEN TRAUMA RECENT>• <ABDOMEN TRAUMA REMOTE HX>• <CHEST PAIN SUBSTERNAL LASTING GTR THAN 20MINUTE <S>>• <CHEST PAIN SUBSTERNAL LASTING LESS THAN 20MINUTE <S>>• No representation of these pairs being “thesame,” but at different times.• Note: Same problem with space; need orthogonality!Motivating Example: Distinguishing FourPossible RelationshipsBetween Transfusion and Jaundice• Post-transfusion antigen incompatibilityhemolytic anemia• Post-Transfusion Hepatitis B: Acute Hepatitis• G-6-PD hemolytic anemia treated by transfusion• Post-Transfusion Hepatitis B: Chronic Activebloodtransfusionabdominalpainjaundice??Even Simple Models Help• PIP’s temporal model:– PAST, RECENT, NOW, SOON, FUTURE• Example:– in “chronic glomerulonephritis” model, “past acuteGN”– (to my surprise), program hypothesized “futurechronic GN” after diagnosing “now acute GN”AGNCGNpast present futurepast present future2What is Time?• (Macroscopically) unidirectional• Related to causality• May be modeled in various ways– Continuous quantity, as in differentialequations– Discrete time points, as in discrete eventsimulations– Intervals, as in ordinary descriptions ofdurations, processes, etc.… or combinationsContinuous View• Differential equation view of world• States (state variables) evolve accordingto their laws002022xtvgtxvgtdtdxgdtxd++=+==Approaches that ExploitContinuous View• Parameter Estimation for a known model– Time-varying parameters• E.g., blood volume, cardiac output, peripheralvascular resistance• Model Estimation– What do you know if I observe a time-delayedresponse to an intervention?Discrete Event View• Designated (countable) time points• Nothing “interesting” between events• Events may be defined by– Clock “ticks”– Interactions among objects in universe– Distinguished points in representation of statevariables (e.g., highest point of cannon shell)What Can Be a Time Point?• Calendrical point—a specific date/time• Recognizable event—e.g., “when I had mytonsils out,” or “start of high school,” or“my ninth birthday”• Now—special, because it movesDiscrete Events areAssociated with State TransitionsE.g., Beck & Pauker’s model to help compute quality-adjusted years of survival:3Constraint Propagationamong Time Points• Clearly, T(A,B)+T(B,C) = T(A,C)• But we only know lower/upper bounds on T• L(A,C) = L(A,B)+L(B,C)• U(A,C) = U(A,B)+U(B,C)• and thus, we can infer relationshipsA CBl, ul, ul, uInitial Assertions• Completingall therelations notexplicitlyassertedANOREXIA>Externally asserted<ANOREXIA<IRRITABILITY3 DAYS2 DAYS7 DAYS5 DAYSLegendInferred3 DAYS4 DAYSConstraint<ANOREXIA<IRRITABILITY ANOREXIA>5 DAYS5 DAYS2 DAYS3 DAYS3 DAYS3 DAYSPropagated Constraint• Order n2 edges in fully interconnected graph• Order n3 computation• Work to localize propagation to semanticallyrelated events<ANOREXIA<IRRITABILITYANOREXIA>5 DAYS5 DAYS3 DAYS2 DAYS2 DAYS3 DAYSIntervals and Points areAlternate Representations• “Overlaps” defined in terms of itsendpoints:<Irritability Irritability><Anorexia Anorexia>+0,++0,++_,+ +_,++_,++_,+Forms of Temporal Uncertainty• Lower/upperbounds on temporaldistances• Central range +fringe• Continuousdistributions4Interval View• Activities, processes take place overextended intervals of time• Observations are true over periods of time• Systems remain in steady (from someviewpoints) states over intervalsAllen’s Temporal IntervalsInference among Intervalsby CompositionX-YY-Ze.g., if A starts B and B overlaps C,what are the possible relationshipsbetween A and C?1. A before C2. A meets C3. A overlaps CInterpreting the Pastwith a Causal/Temporal ModelInterpreting the Pastwith a Causal/Temporal Modelweakheartheartfailuredigitaliseffectretainlosediureticeffecthighlowedemafluid therapywaterbloodvolumelowcardiacoutputdefinite causepossible causepossible correction (not all shown)PostdictionPostdictionweakheartheartfailuredigitaliseffectretainlosediureticeffecthighlowedemafluid therapywaterbloodvolumelowcardiacoutputdefinite causepossible causepossible correction (not all shown)Long, Reasoning about Statefrom Causation and Time in aMedical Domain, AAAI 830345678910nowfuturepastpresentnormhigh ? norm lowretain? loss ?lowpresentpresentpresentpresentedemablood volumewatercardiac outputheart failureweak heartdiuretic effectdiuretic12Temporal Control Structure.- T. Russ• Processes maintaining truth of abstractions overspecified interval• Update of past beliefs from corrections or new data.• Actions are permanent.5Back to Monitoring• Detecting Trends• Language for Trend Description• Matching algorithms• Top-down vs. bottom-up vs. both• Learning trend detectors“Two-Point” Trend Detectors• Restricted to hospitals with the mostcomplete information systems• Rind & Safran, 1992• Two point event detector– rise in creatinine > 0.5 mg/dl• Therapeutic context– Renally cleared or nephrotoxic medication– Possible care providers• Implementation– M procedures linked to E-mailBIH Experience (cont’d)• Time series trial• 607 in 348admissions duringcontrol periods• 497 events inintervention period• 369 alerts, sent to584 differentphysicians, 9.25recipients per alert• Improved responsetime• Improved outcomesRepresentationEasily implemented as an Arden SyntaxMLMAre these two-point trenddetectors sufficient?• If not, why not?• Noisy data.• Multi-phased processes.• Uncertainty over time.• Uncertainty over values.Issues in Trend Detection• Defining significant trends– Multiple variables– Multiple phases– Temporal and value uncertainty• Detecting trends from data• Generating alarms• Displaying, explaining results• Changing clinical contextPediatric Growth MonitoringPediatric Growth Monitoring• Data:- heights, weights- family history- bone ages- pubertal data,stages- hormone values• Disorders