Η λίστα ιστολογίων μου

Related Posts

Δημοφιλείς αναρτήσεις

Medicine by Alexandros G.Sfakianakis

Medicine by Alexandros G.Sfakianakis,Anapafseos 5 Agios Nikolao

Medicine by Alexandros G.Sfakianakis

OtoRhinoLaryngology - Head and Neck Surgery

Εμφανιζόμενη ανάρτηση

Peripheral positional vertigo and dizziness (PPVD)

The diagnostic framework of peripheral positional vertigo and dizziness (PPVD): a new concept based on the observation of alcohol-induced po...

Τρίτη, 29 Νοεμβρίου 2016

Finding Factors Associated With Post–Emergency Department Morbidity and Mortality in Elderly Patients: Analyzing a Case-Control Study

Finding Factors Associated With Post–Emergency Department Morbidity and Mortality in Elderly Patients: Analyzing a Case-Control Study: Q1.a The accurate identification of predictors of patient morbidity and mortality may be critically important to patient management. However, these predictors are often challenging to reliably ascertain. Describe some of these challenges that are applicable to all emergency department (ED) patients. Discuss the unique challenges in the elderly ED patient population.

Discussion Points

  • 1.
    Inline Image fx2A. The accurate identification of predictors associated with patient morbidity and mortality may be critically important to patient management. However, these predictors are often challenging to reliably ascertain. Describe some of these challenges that are applicable to all emergency department (ED) patients. Discuss the unique challenges in the elderly ED patient population.
    • B.
      Inline Image fx1The primary decisions of study population and outcomes are suitable for a number of reasons. Why did the authors limit the study population to elderly patients who were discharged from the ED? What could have been confounders in this population of the primary outcomes: “combined poor outcome of either death or an ICU admission shortly after ED discharge”?1
    • C.
      Inline Image fx2This study was a case-control design. Why might investigators choose a case-control study design instead of a cohort study? Name at least 2 key characteristics of the problem to be investigated that play into this decision. Describe the advantages and disadvantages of a case-control study compared with a cohort study. How does the investigator determine the appropriate number of controls for each case?
    • D.
      Inline Image fx2How might you have designed a cohort study that used existing data to investigate the primary study outcome? Would you have imposed the same inclusion and exclusion criteria on the study population? What if you conducted a cohort study in which you collected new data?
  • 2.
    Inline Image fx1A. This study examined a set of predetermined variables and measured whether they were associated with the composite outcome of death or ICU admission shortly after ED discharge. Why might investigators identify potentially important clinical predictors a priori versus relying on regression analyses to identify predictors? Might there be disadvantages to predetermined potential predictors?
    • B.
      Inline Image fx2Clinical variables (eg, age, systolic blood pressure) may be recorded and analyzed as continuous, ordinal, or binary (ie, dichotomous) terms. Define each of these terms and provide an example for each, using patient age. Variables in this article were treated as binary rather than continuous for the purposes of analysis. What are the pros and cons of converting continuous variables to binary terms?
    • C.
      Inline Image fx3A study’s design can increase or decrease the possibility of confounding (bias). Please comment on how each of the following factors might affect the likelihood of confounding.
      • i.
        use of paired controls
      • ii.
        exclusion of do not resuscitate patients, do not intubate patients, and hospice patients
      • iii.
        patient selection (consider demographics, socioeconomics, etc)
      • iv.
        selection of periods (patient data period, poor outcome period)
      • v.
        variable selection
      • vi.
        chart abstractor training and work processes
      • vii.
        modeling
      • viii.
        inclusion of against medical advice patients in “change of disposition” parameter
    • D.
      Inline Image fx1In this article, the quality of the data relies heavily on the quality of the assessments of the chart abstractors. For all assessments conducted by chart abstractors, what else could be done to mitigate effects of human error on the quality of analysis?
  • 3.
    Inline Image fx1A. What other variables (eg, patient lives alone, history of dementia) would you have included in the study analysis? How would you define those variables (eg, normal with X range of values, abnormal with Y range)? Consider if the model included a patient’s social support system. What would you expect the association between that and the combined poor patient outcome to be?
    • B.
      Inline Image fx1In the selected model, variables related to consultants were left out. Explain this decision by the authors. Why is this justifiable? Consider the overall goal of the article.
    • C.
      Inline Image fx1Will this study change clinical practice at your hospital? Support your opinion.

Answer 1

Q1.a The accurate identification of predictors of patient morbidity and mortality may be critically important to patient management. However, these predictors are often challenging to reliably ascertain. Describe some of these challenges that are applicable to all emergency department (ED) patients. Discuss the unique challenges in the elderly ED patient population.
All ED patients:
  • a.
    Some factors are easy to identify but difficult to quantify, such as social factors, psychological temperament, level of faith, general intelligence, and desire to live.
  • b.
    Some factors are easy to identify and quantify but difficult to gather, such as lifetime tobacco or radiation exposure.
In addition to the above, the elderly population has additional barriers in achieving an accurate set of factors in this predictive model:
  • a.
    cognitive capacity of patient to provide accurate data
  • b.
    unwillingness to engage through modern research conduits such as smartphones, computer-administered surveys, or automated telephonic surveys
  • c.
    complexity of each individual, given total years lived
Q1.b Why did the authors limit the study population to elderly patients who were discharged from the ED? What could have been confounders in this population of the primary outcomes: “combined poor outcome of either death or an ICU admission shortly after ED discharge”?1
The study population was ideal for 2 reasons. First, the elderly population likely has the highest yield in the chosen outcome irrespective of whether they were discharged from an ED or not. Second, although a younger population could have been included, this would likely weaken the association of the explanatory factors and the outcome because different factors play a role in the outcome of death or ICU admission after ED discharge in a younger population.
The study could have been confounded by factors that led to discharge not because the physicians wrongly assumed that the patient was at low risk for deterioration but because a decision was made to discharge the patient despite a high risk of deterioration. This could occur if a patient were receiving hospice care or were thought to have a better chance at home than in the hospital even though outlook was poor for both choices. The presence of such patients in the group scored as “having the outcome” might confound a study designed to predict what factors are associated with unanticipated poor outcome.
Q1.c This study was a case-control design. Why might investigators choose a case-control study design instead of a cohort study? Name at least 2 key characteristics of the problem to be investigated that play into this decision. Describe the advantages and disadvantages of a case-control study compared with a cohort study. How does the investigator determine the appropriate number of controls for each case?
Investigators may choose a case-control study design when a disease is rare or there is a long interval between the exposure and development of the disease. The financial and logistic requirements of a cohort study make that study design not feasible to conduct for rare conditions because of the massive number of patients that would be needed to capture the required number of patients with the disease. The same limitations apply when the development of a disease is investigated many years after an exposure (eg, investigating whether exposure to an artificial sweetener as a teenager is associated with increased risk of solid organ tumors after aged 50 years).
Case-control studies provide a structured, although less methodologically robust, study design for studying such conditions. Readers interested in a more detailed discussion on this topic are directed to the July 2013Annals Journal Club answers.2 Case-control studies are useful for efficiently investigating rare outcomes and generating hypotheses that may be further studied in more rigorous investigations.3 The main threat to the internal validity of case-control studies is recall bias, that case patients and controls have different levels of motivation or different capacities for accurately remembering their exposure status. The other major threat is bias arising from the case patients’ and controls’ being selected separately and the investigators having insufficient means of controlling for differences between them. Case-control studies investigate a single outcome or disease, whereas cohort studies may look at multiple outcomes.2, 3
Identifying controls is generally the easier part of a case-control study, and determining the number of controls depends on the statistical power calculation for the investigation. In general, optimal precision is obtained when the ratio of controls to case patients is 2:1 or 3:1. Little precision is gained if more controls are added.2
Q1.d How might you have designed a cohort study that used existing data to investigate the primary study outcome? Would you have imposed the same inclusion and exclusion criteria on the study population? What if you conducted a cohort study in which you collected new data?
In a cohort study, the groups are created according to the exposure (in this case, being discharged from the ED) and then followed for the outcome (death or admission to the ICU). This could be conducted with any database that had the relevant variables on a complete set of ED patients. The inclusion criteria would have to be tailored to the specific question being considered. For example, the reasons that a group of frail elderly fare poorly after discharge might be very different from the reasons a group of vital elderly or middle-aged persons fare poorly. The main problem with a cohort study of this kind performed on existing data is that certain key variables may not be consistently available.
As implied by the previous answer, the real issue here is not who fares poorly but who fares poorly when it was anticipated their outcome would be good. It is unlikely that a cohort study of existing data will ever be able to answer this question. Only in a prospective study in which the physician could be asked how he or she expected the patient to fare could this be achieved.

Answer 2

Q2.a This study examined a set of predetermined variables and measured whether they were associated with the composite outcome of death or ICU admission shortly after ED discharge. Why might investigators identify potentially important clinical predictors a priori versus relying on regression analyses to identify predictors? Might there be disadvantages to predetermined potential predictors?
Choosing variables for a model is a complex task, and much has been written about both theoretical and pragmatic aspects of model selection and specification.4, 5, 6 Although there is no absolute proof that any particular perspective is correct, the Annals of Emergency Medicine editorial board generally believes that variables chosen for models should be based on existing evidence, theory, and reason rather than the complete set of all possible variables that one could put into a regression.
As discussed many times in this series, statistical significance confounds sample size and effect size. As a consequence, an important variable (eg, undiagnosed rabies) might not be included if it is highly predictive but few persons have it, whereas in a large data set a variable that has much less discriminative power but commonly occurs may enter the data set. Of course, using statistical selection criteria for variable selection guarantees that some variables will be entered as a result of type I error (based on the α selected) and will enter the model despite having no true predictive value (overfitting).6
The other side of this argument is that one may only include variables that one knows about. Shotgun approaches to identifying causative variables will periodically identify new candidate predictors, and some of these may turn out to be valid.
One approach to having the best of both worlds is to have exploratory models that may use shotgun approaches while requiring that any attempts at deriving or validating prediction models use a theoretically based set of variables.4, 5, 6
Q2.b Clinical variables (eg, age, systolic blood pressure) may be recorded and analyzed as continuous, ordinal, or binary (ie, dichotomous). Define each of these terms and provide an example for each, using patient age. Variables in this article were treated as binary rather than continuous for the purposes of analysis. What are the pros and cons of converting continuous variables to binary terms?
  • • Continuous variable: one that has an infinite number of possible values
    ∗.In mathematics, a continuous variable’s domain should extend from –∞ to ∞. For our purposes, we consider a variable continuous if its domain is sufficient compared with the values considered in the study so that truncation is not an issue. For example, in a study of middle-aged persons we would not be terribly concerned that age is continuous only between 0 and 110 years or so. Alternatively, in a study of centenarians, age should not be treated as a continuous variable for statistical purposes because there will be patients whose age is at the domain’s boundary.
    • ○ Example continuous ages: 0.3, 12.2, and 51.5
  • • Ordinal variable: one that is ranked but the rankings do not imply a specific distance between ranks, as in first, second, third
    • ○ Example of ordinal ages: first child, second child, third child
  • • Categorical variable: One that takes on some finite number of discrete categories with no order to the categories
    • ○ Example: patient race
  • • Binary variable: a categorical variable that takes on one of 2 states (eg, on/off)
    • ○ Example of binary ages: younger than 65 or older than or equal to age 65
Continuous variables may be converted to binary variables as above. This is always achieved with some risk, as illustrated in the Figure. If the situation is as shown in Figure A, then little information is lost by making 2 categories based on the location of the green line. However, if the situation is more like that shown in FigureB, then there is no obvious cut point because one signified by the red line would combine the low values of the youngest persons with the middle values of middle-aged persons, whereas a cut point at the blue line would average the results for middle-aged and oldest persons. Obviously, the consequences of binary collapse are even more deceptive when the relationship is not monotonic.
Thumbnail image of Figure. Opens large image

Figure

Examples of distributions that are conducive to conversion from continuous to binary (A) and inconducive (B).
The advantage of using binary variables in predictive models is that it makes the model easier to implement. The model is easier to use if it uses 1 point for being old (0 points for being not old) than if it uses “add 0.05 points for each year of age above 30 years.”
Q2.c A study’s design can increase or decrease the possibility of confounding (bias). Please comment on how each of the following factors might affect the likelihood of confounding.
Confounding (bias) is unavoidable. Investigators can choose strategies that attempt to minimize it but, in clinical research, it is impossible to conduct a study that is not vulnerable to some bias.
  • 1)
    Use of paired controls: By pairing study cases with control cases, one hopes to decrease confounding by eliminating as many between-group differences as possible.
  • 2)
    Exclusion of do-not-resuscitate, do-not-intubate, and hospice patients: As discussed above, the goal is to determine factors associated with unanticipated deaths. The exclusion of those for whom death may be expected helps focus the study on patients who are expected to fare well. The failure to exclude such patients (for example, if there were no documentation that the patient was receiving hospice care) would make estimates of unexpected deaths too high and might alter the associations with predictor variables.
  • 3)
    Patient selection: Patients were selected according to multiple criteria, including age, disposition, and existence of hospice or advanced directives. Potential selection criteria that were not already mentioned in this journal club include:
    • a)
      Member of Kaiser Permanente health plan: Kaiser Permanente patients may represent a different demographic or socioeconomic status than the general population. They also likely have a more robust system of follow-up, and that may influence the emergency physician’s decision toward more discharges than at other EDs. These factors may limit external validity of the study.
  • 4)
    Selection of periods
    • a)
      One-year enrollment window: By using a 1-year window, the authors reduce the possibility of temporal (seasonal) confounding.
    • b)
      Seven-day outcome window: A 7-day period is reasonable because longer intervals may capture deaths that had nothing to do with ED care and shorter intervals might miss those that are attributable, at least in part, to ED care.
  • 5)
    Variable selection: It appears that variable selection was based on theory because “all measures were included in a full conditional logistic model.” This is a better strategy than selecting variables by their univariate statistical significance.
  • 6)
    Chart abstractor training and work processes: Abstractors were trained, tested on their training, and blinded to the study goals. They also had an open line of communication with investigators to resolve disputes. Because they were blinded, the abstractors had no way to be biased and also had no obvious incentive to bias the work.
  • 7)
    Modeling: As with any retrospective review, one can model only those variables for which data can be collected accurately from the chart. This is a limitation of all modeling. Any modeling effort makes assumptions both in the model form and the model specification.4 The choices here seem reasonable; however, one might wonder, as discussed elsewhere in these answers, whether the variables collected are sufficient to differentiate expected from unexpected deaths.
  • 8)
    Inclusion of ED patients who leave against medical advice in “change of disposition” parameter: The issue is not so much whether the patients who leave against medical advice should be included but whether they should be grouped with patients whose disposition was changed by their physician. There is no strong reason to believe that the association between patients who leave against medical advice and the outcome and the association of patients whose disposition was changed and the outcome should be the same. Given this, it is impossible to know whether combining these confounds the analysis.
Q2.d In this article, the quality of the data relies heavily on the quality of the chart abstraction. For all assessments conducted by chart abstractors, what else could be done to mitigate effects of human error on the quality of analysis?
Data abstraction is one of the most critical elements of any retrospective analysis of written records (be they handwritten or computerized), and there are many outstanding references on proper methodologies to minimize potential bias. Two such articles were authored by Gilbert et al7 and Kaji et al.8 We highly recommend them for readers interested in this topic. Investigators who plan to perform any chart review study should read these articles and follow the recommended methodology when conducting data abstraction and recording.7, 8

Answer 3

Q3.a What other variables (eg, patient lives alone, history of dementia) would you have included in the study analysis? How would you define those variables (eg, normal with X range of values, abnormal with Y range)? Consider if the model included a patient’s social support system. What would you expect the association between that and the combined poor patient outcome to be?
Any number of variables could have been included in this analysis. For a support system, one might define a range of possible support tools such as home medical alert systems, home health aids, scheduled family or friend visits, or skilled nursing or assisted living facilities. Investigators might assign dichotomous (0/1) scores to each of these social support tools or develop a continuous scale grading these from no support (0 points) to some support (eg, family visit weekly, 5 points) to most intense (skilled nursing facility, 20 points). One might opine that a lack of social support is associated with a higher level of combined poor outcome after ED discharge. Developing and validating such a scale would be a project as big as or bigger than the one reported but would be necessary unless such a measure already existed.
Q3.b In the selected model, variables related to consultants were left out. Explain this decision by the authors. Considering the overall goal of the article, is this justifiable?
The authors chose to exclude specialty consultation from the model after their analysis found that specialty consultation and “change in disposition plan” agreed in 88% of cases, suggesting that involving a consultant frequently resulted in a change in disposition plan. The authors likely thought it was more plausible that a change in disposition plan from admit to discharge was associated with a poor outcome than the involvement of a consultant. Because the goal of the study was to identify factors that could help physicians prevent poor outcomes on discharging the elderly, it is also easier to assimilate the idea of “change of disposition” than “obtaining a specialty consultation” into a plausible warning signal for the provider.
Q3.c Will this study change clinical practice at your hospital? Support your opinion.
The frequency of ED visits by elderly patients will continue to increase with the aging of our population. It is neither practical nor responsible to admit every elderly patient according to their age. However, specific characteristics are associated with worse short-term outcomes for this patient population. Emergency physicians and admitting hospitalists should continue to remain vigilant for the vital sign abnormalities and cognitive deficits that can signify danger. Future investigations with geriatric ED patients should provide additional insight into the relative importance of social factors and physician gestalt compared to the factors explored in this effort.

References

  1. Gabayan, G.Z., Gould, M.K., Weiss, R.E. et al.Poor outcomes following emergency department discharge of the elderly: a case-control study. Ann Emerg Med20166843–51
  2. Gupta, M., Barrett, T.W., and Schriger, D.L. Every peddler praises his own needle: have clinical rules in the diagnosis of subarachnoid hemorrhage supplanted lumbar punctures yet?. Ann Emerg Med2013;62633–640
  3. Hulley, S., Cummings, S., Browner, W. et al. Designing Clinical Research. 3rd ed. Lippincott Williams & WilkinsPhiladelphia, PA2007112–126
  4. Greenland, S. Modeling and variable selection in epidemiologic analysis. Am J Public Health1989;79340–349
  5. Schriger, D.L. Suggestions for improving the reporting of clinical research: the role of narrative. Ann Emerg Med200545437–443
  6. in: F.E. Harrell Jr. (Ed.) Regression Model Strategies, With Applications to Linear Models, Logistic Regression, and Survival AnalysisSpringerNew York, NY2001
  7. Gilbert, E.H., Lowenstein, S.R., Koziol-McLain, J. et al. Chart reviews in emergency medicine research: where are the methods?. Ann Emerg Med199627305–308
  8. Kaji, A.H., Schriger, D.L., and Green, S. Looking through the retrospectoscope: reducing bias in emergency medicine chart review studies. Ann Emerg Med201464292–298
Editor’s Note: You are reading the 52nd installment of Annals of Emergency Medicine Journal Club. This Journal Club refers to the article by Gabayan et al1 published in the July 2016 edition of Annals. Information about journal club can be found at . Readers should recognize that these are suggested answers. We hope they are accurate; we know that they are not comprehensive. There are many other points that could be made about these questions or about the article in general. Questions are rated “novice” (Inline Image fx1), “intermediate” (Inline Image fx2), and “advanced (Inline Image fx3) so that individuals planning a journal club can assign the right question to the right student. The “novice” rating does not imply that a novice should be able to spontaneously answer the question. “Novice” means we expect that someone with little background should be able to do a bit of reading, formulate an answer, and teach the material to others. Intermediate and advanced questions also will likely require some reading and research, and that reading will be sufficiently difficult that some background in clinical epidemiology will be helpful in understanding the reading and concepts. We are interested in receiving feedback about this feature. Please e-mail with your comments.

Δεν υπάρχουν σχόλια:

Δημοσίευση σχολίου