- Email: [email protected]
Does prediction of outcome alter patient management?
Saturday
12
June 1993
No 8859
ARTICLES
Does prediction of outcome alter patient
management?
A patient’s prognosis is a key factor for the clinicians involved in management. We set out to determine if provision of computer-based predictions of outcome after severe head injury resulted in measurable changes in patient management. In particular, we wondered whether introduction of the predictive system would alter the relation between severity of injury and "intensity" of
management.
patients admitted to four British neurosurgical units between 1986 and 1989 following a severe head injury, and who were either in coma for 6 h or had an operation for acute intracranial haematoma, were studied. Specified 1025
recorded and all patients were followed up after six months. The study had three phases: a baseline period of at least one year before the introduction of computerbased outcome prediction, one year when predictions were provided at specified times, and a final six months when prediction was withdrawn. While predictions were being provided, there was an increase in the use of specified aspects of intensive care in patients predicted to have a good outcome, but a 39% reduction in the use of these same aspects of intensive care in patients predicted to have the worst outcome. There was no evidence that the provision of predictions affected overall outcome, length of stay, or the recording of explicit decisions to limit treatment. We have demonstrated that the introduction of a routine prediction service can alter patient
aspects of intensive management
management.
were
Introduction The process by which a doctor arrives at a management decision for a seriously ill patient is often complex. Many factors have to be taken into consideration and balanced against one another. Information about the patient’s clinical state and items such as age and the results of biochemical
investigations
or
imaging procedures
are
important.
Resource availability is also relevant. For humanitarian, ethical, and economic reasons doctors are being encouraged to concentrate activities and resources on patients who are most likely to gain benefit, and to minimise intensive treatment for patients whose suffering is thereby prolonged or who would be left with an unacceptable quality of life. Conversely, undue pessimism, often based on a single criterion such as age, can lead to premature limitation of treatment, whereas unnecessary or unduly extended intensive care can expose patients with less severe injury to risks.l The expected eventual outcome is a crucial factor in making decisions about a patient’s management.2 Physicians’ estimates of outcome are often unduly optimistic or pessimistic or unnecessarily ambiguous,2-5 and a less subjective approach to the problem of predicting a patient’s future is required. Methods of predicting outcome in ADDRESSES: Department of Neurosurgery, University of Glasgow, Southern General Hospital, Glasgow G51 4TF, UK (L. S. Murray, PhD, Prof G. M. Teasdale, FRCS, Prof B. Jennett, MD, D. Kelly, BN); Department of Surgery, University of Glasgow, Western Infirmary, Glasgow (G. D. Murray, PhD); Department of Clinical Neurosciences, University of Edinburgh, Western General Hospital, Edinburgh (Prof J. D. Miller, MD, P. Jones, MAppSc); Wessex Neurological Centre, University of Southampton, Southampton General Hospital, Southampton (Prof J. D. Pickard, FRCS, S. Bailey, SRN); and The Walton Centre for Neurology and Neurosurgery NHS Trust, Walton Hospital, Liverpool (M. D. M. Shaw, FRCS, J. Achilles, RGN, J. Lacey, RGN). Correspondence to Dr L S. Murray.
1488
TABLE I-FEATURES OF PATIENTS IN DIFFERENT STUDY PERIODS ON ADMISSION TO NEUROSURGICAL UNITS
critically ill patients have been developed, but the effect on patient management of providing clinicians with this predictive information has not been widely examined. Knaus et al9 described changes in the pattern of treatment decisions that resulted from provision of information about the chances of survival of patients with multiple organ failure. Severe head injuries account for a substantial part of the work of neurosurgical units and many general intensive-care units. In many neurosurgical patients, early treatment efforts are rewarded by reasonable recovery, but even with the most intensive regimens of monitoring and therapy mortality for head-injured patients in coma is 30-40% and another 15-20% survive with persistent severe disability. We report the effect on patient management and outcome of providing clinicians with estimates of prognosis after severe head injury.
Patients and methods Patients The four centres that participated in the study-Glasgow, Edinburgh, Liverpool, and Southampton-provide a range of neurosurgical facilities for head-injured patients that embraces the
patterns of care in most UK regions. The patients studied were either in coma (ie, no eye opening, not speaking, and not obeying commands) for more than 6 h or had an acute intracranial haematoma that needed to be evacuated by craniotomy. Research nurses collected information on the clinical state of each patient throughout their stay, the investigations carried out, treatments given, the duration of treatments, and length of stay. Outcome was assessed at six months after injury according to the Glasgow outcome scale10 with three outcome groups: dead/ vegetative state, severe disability, and moderate disability/good recovery.
Assessment of patient management
Management was characterised by features that were deemed, on the basis of clinical judgment, to be those most likely to be influenced by a clinician’s perception of outcome. The use of invasive investigation (intracranial pressure monitoring), intensive treatment (intubation and ventilation, osmotic diuretics), and written instructions to limit treatment were noted. The time till death or, in survivors, the length of stay was also calculated. In addition to comparing the overall management practices, we postulated in advance that management might be influenced differently in cases with different severities of injury.
Experimental design The design adopted has been characterised as quasiexperimental." Thus, the study, which took place between 1986 and 1989, was divided into three periods: a baseline period of at least one year during which data were collected but no predictions of outcome given to staff, one year when the prediction program was used, and a final six months when prediction was withdrawn. Findings from the three periods were then compared. This approach was adopted because we believed that it was relevant to assess the overall performance of a neurosurgical department, and that randomisation of patients or clinicians or of the small number of centres would not be appropriate. The introduction of the prediction service was staggered at the four centres to take account of coincidental changes in practice due to other factors.
Before the system was introduced into each centre,
validity was established by comparing the predicted and actual outcome in patients who had been followed up. When the prediction service was introduced, the principles of calculating prognosis and the interpretation of probabilities were explained and the validity of the system demonstrated to staff at each centre. The observational nature of the study was emphasised and care was taken not to specify how predictions should be used to change practice. During the period when prediction was available, research nurses continued to collect patient data and also made predictions on the lst, 3rd, and 7th days after onset of coma, or before operation for those patients with an intracranial haematoma. Results of prediction were provided as a regular service, in a way similar to a laboratory report. Clinicians were made aware of results during clinical rounds. Medical and nursing staff had free access to the computer and were able to make predictions. If predictions were not made by ward staff at weekends they were provided by the research nurse as soon as possible afterwards. The prediction service was stopped after one year and the its
computer removed. Collection of data about new cases continued for a further six months and all cases were followed up for six months.
Statistical methods Outcome prediction The computer program that was used to calculate predicted outcome was designed so that it could be used by neurosurgical and nursing staff without special training, and run on a personal computer (Amstrad PCW8256). The program used an independence model based on Bayes’ theorem7,11 to update the probability of each outcome given the clinical information about the patient. First, the program requested patient and user identification and displayed appropriate prior probabilities of outcome. The user was then questioned about the patient’s age, findings of the Glasgow coma scale, changes in clinical condition, motor-response pattern, pupil reaction, and eye signs. After the entry of each item of clinical information, updated probabilities of outcome were calculated and displayed. Results were also presented by two graphical methods and printed copies provided. The decision to base predictions on the independence model was a pragmatic one, in that it was simple to make the predictions even when some prognostic variables were unavailable. This approach yielded predictions that were similar to those obtained by more sophisticated statistical modelling, the only disadvantage being that the predictions tended to be less well calibrated.’2
The effect on outcome of providing predictions was assessed by a logistic regression model that related the probability of a poor outcome (dead or vegetative) to the phase of study (before predictions vs during predictions), after correcting for the effects of age, pupil reaction, and Glasgow coma scale at admission to the neurological unit. Such a model has been shown previously, with similar
data,
to
give
an
excellent fit for the data, and
to
allow
development of a powerful, well calibrated scoring system for predicting outcome.14 The effect on patient management of providing outcome predictions was assessed first by dividing the patients into three prognostic groups, based on the logistic regression model, that were defined by the estimated probability of a poor outcome being 00 to 04 (good prognosis), 04 to 08 (moderate prognosis), or 08 to 10 (poor prognosis). The proportion of patients receiving various aspects of intensive management was then related to the prognostic group and to the phase of the study (before predictions vs during predictions) with a standard log-linear model. Particular emphasis was placed on the "interaction" term, which reflects any change in the relation between prognosis and management that was associated with the introduction of the prediction service.
1489
TABLE 11-OUTCOME SIX MONTHS AFTER INJURY IN DIFFERENT STUDY PERIODS
Percentages are shown unless otherwise
TABLE IV-USE OF INTENSIVE MONITORING AND TREATMENT IN DIFFERENT PROGNOSTIC GROUPS BY STUDY PERIOD
indicated.
These formal analyses did not incorporate the data from the final phase of the study, after the prediction service had been withdrawn. This was because the primary hypothesis of the study related to the effects of introducing the predictions, and, moreover, given the limited sample size for the final phase, any formal analysis of these
data would lack statistical power. However, data from the final
phase are reported to allow informal comparisons to be made. Results
patients were studied-38% in Glasgow, 25% in Edinburgh, 23% in Liverpool, and 14% in Southampton. The proportions of patients from the different centres and the types of patients admitted and their characteristics were similar in the different periods of the study (table l). In some patients it was not possible to make an outcome prediction because the use of paralysis and ventilation eliminated the physical signs necessary for the predictive system. This was the case in 1%, 19%, 16%, and 13% of patients pre-operatively, at 24 h, at 3 days, and at 7 days, respectively, during the period of the study. During the period when predictions were provided to medical staff, outcome prediction was available in 94% of appropriate 1025
cases.
6 months after injury in the different is shown in table II. Overall, just over study periods one-third of the patients died and almost half had a moderate disability or made a good recovery. The logistic regression model, which corrected for age and severity of injury, indicated no difference in outcome between the periods before and during which predictions were provided (p = 0-57, X2). This lack of effect of prediction on outcome was consistent over the three prognostic groups, with no substantial differences between the observed and expected numbers of poor outcomes in the three groups. There was an overall trend for a modest, but non significant, increase in intracranial-pressure monitoring, intubation and/or ventilation, and the administration of osmotic diuretics during the period when predictions were provided (table III). However, analysis of the interaction between initial severity and management showed variations in the different phases of the study. Thus, when predictions
Patient
outcome
TABLE III-USE OF ASPECTS OF MANAGEMENT IN DIFFERENT STUDY PERIODS I
I
I
*Patients are stratified ie, dead or vegetative.
according to their estimated probability of a
poor outcome-
provided, the pattern of use of intracranial-pressure monitoring, intubation and/or ventilation, and osmotic diuretics differed substantially compared with the period before prediction in patients in the same estimated prognosis group (table IV). Among patients whose features indicated a moderate or good prognosis, osmotics, intubation and/or ventilation, and intracranial-pressure monitoring were used more frequently during the period that predictions were provided. Conversely, during this period, there was a decrease of 39% in the average frequency of use of these aspects of intensive management in patients with the worst prognosis. This change in the relation between prognosis and management was statistically significant (X2 test) for the use of osmotics (p <0 01) and intubation and/or ventilation (p < 0-001). The pattern of redeployment of osmotics and intubation and/or ventilation away from patients with a poor prognosis and towards patients with a moderate or good prognosis was seen consistently across the four participating were
centres.
A decision to limit treatment was recorded in the case record of about 1 in 5 patients; this finding did not vary between the study periods, and was followed by a death or a dependent outcome in more than 9 out of 10 patients (95%, 98%, and 92% before, during, and after predictions, respectively). Also, there was no variation between the three periods in the time to death in patients with a fatal outcome or in the duration of stay in the neurosurgical unit of survivors (table III).
Discussion The development and evaluation of a mathematically based clinical decision aid can be likened to the phases in assessment of a new therapy. 13 In phase I the model is developed, in this case the independence model for predicting outcome in the early stage after severe head injury, based on easily obtained clinical signs.’1 In phase II, the quality of the information provided is assessed by comparing the results obtained from the predictive model with those of other statistical models and with eventual clinical outcome. We have reported previously that when clincial features are well selected, the independence model performed as well as, or even better than, more complex techniqueslz and gave predictions that were more consistent
1490
and reliable than those made by senior neurosurgeons.14 Our study represents the third phase of the development and evaluation process. It deals with the discovery of the practical consequences of providing staff in neurosurgical units with estimates of prognosis for severely head-injured patients. The system proved acceptable to staff, and its provision was associated with redeployment of some aspects of intensive care from patients with a poor prognosis to those with a better estimated prognosis. We assessed the effects of predictions by studying series of patients in different units in successive periods, rather than focusing on the treatment of individual patients or the actions of individual doctors, and we rejected the possibility of randomising patients for whom prognostic information would be provided. Neurosurgical units usually function with considerable cross-cover by junior doctors, particularly in emergency situations, so that a patient’s management is influenced by the unit’s general attitudes and policies. The mere presence of the predictive system in a unit could, therefore, have a confounding effect on the management of patients for whom predictions were not made. For similar reasons, the consequences of predictions on specific patients could not be separated from the general effect of the use of the system. The value of correction of hypotension and hypoxia, of the prompt evacuation of intracranial haematomas, and of the general intensive care of the unconscious head-injured patient is universally accepted. However, there is less agreement about the benefits of other components of intensive management of head injuries, and this is reflected in the variations in practice shown in this study. Thus, in different centres the proportion of patients treated with osmotic diuretics varied from 23% to 67%, with intubation and/or ventilation from 44% to 83%, and intracranialpressure monitoring from 1% to 57%. This finding is in accord with other reports15.16 of considerable variations in the use of these and other aspects of intensive management. When there are no clearly established indications for the use of a component of treatment, its employment may be governed by an explicit protocol or, more often, may be determined by specific decisions in each case. Our hypothesis that prognosis would be a factor in such decisions was confirmed by our findings. Thus, during the period when predictions were provided, osmotic diuretics, intubation and/or ventilation, and perhaps intracranial pressure monitoring were used less often in patients with a very poor prognosis whose outlook was unlikely to be altered, and more often in patients with either an uncertain or optimistic outlook who were more likely to benefit. This change was achieved partly by a slight overall increase in such treatments, but more so by redeployment with less use in cases with the worst prognosis. The absence of an alteration in outcome in association with these changes in treatment indicates that, as used in this study, these components of treatment did not have a crucial effect on outcome.
Our results, which indicate that clinicians respond appropriately to predictive information, should allay concern that an adverse prediction might result in unreasonable pessimism. The frequency with which explicit instructions were given to limit treatment varied from 13%
27% in the different units but was not related to the provision of predictions. Most fatalities occurred within 3 days of injury, and awareness of this early mortality may reduce the pressure on neurosurgical staff to take early decisions to limit treatment. to
Most clinical decision aids that have been subjected to phase III trial-the formal assessment of their influence on clinical practice-have been concerned with diagnosis and
the effects of their use have been variableY-19 However, Knaus et al9 studied patients with failure of three or more organ systems, and found that when objective estimates of a low chance of survival were provided, there was a small but significant increase in decisions to stop active treatment and to continue only with comfort care. The greatest effect was in the most seriously ill patients, in whom "comfort care only" decisions increased from 11 % to 36%, while the total proportion of patients receiving unlimited treatment remained stable. An earlier study20 of bum victims showed a similar effect of prediction, but was criticised as discouraging treatment before the patient’s response to treatment was known.21 Other systems for providing outcome predictions for individual head-injured patients have been devised in Richmond, Virginia, USA,8 and Leeds, UK,22 The Richmond system aims to provide information about the likelihood of different outcomes for a patient, but its influence on practice has not been reported. The Leeds score is based on data obtained within 12 h of admission to the neurosurgical unit, including the results of invasive monitoring, and it aims to identify patients whose death can be predicted with certainty. However, a study from Texas, USA, has shown that the Leeds model cannot be relied on "to clarify the complicated emotional, moral, legal, and financial issues that surround the early termination of care in
seriously head-injured patients" .23 We have examined the effect of providing clinicians with information about the expected outcome after severe head injury, but this approach could have a wider application in patient management in general. We believe that it is not realistic to expect a statistical predictive system to relieve clinical staff of their duty to take account of a range of factors when making decisions about the management of patients. These include concepts such as "error" and "harm" as well as clinical and laboratory data. An estimate of the likelihood of outcome is only one factor for the clinician to consider, and it has been likened to any other test.24 This study indicates that provision of predictive information can result in appropriate measurable changes in clinicial practice, without any evidence of adverse effects on outcome. We thank our neurosurgical colleagues for allowing us to study their patients and for their cooperation with the collection of data and provision of predictions. This study was supported by the Medical Research Council Special Project Grant G842699.
REFERENCES 1.
use of intensive care. BMJ 1984; 289: 1209-11. 2. Barlow P, Teasdale G. Predictions of outcome and the management of severe head injuries: the attitudes of neurosurgeons. Neurosurgery 1986; 19: 989-91. 3. Chang RWS, Lee B, Jacobs S. Accuracy of decisions to withdraw therapy in critically ill patients: clinical judgement versus a computer model. Crit Care Med 1989; 17: 1091-97. 4. Dawes RM, Faust D, Meehl RE. Clinical versus actuarial judgement. Science 1989; 243: 1668-74. 5. Kaufmann MA, Buchmann B, Scheidegger D, Gratzl O, Radu EW. Severe head injury: should expected outcome influence resuscitation and first-day decisions? Resuscitation 1992; 23: 199-206. 6. Knaus WA, Draper EA, Wagner DP, Zimmerman JE. APACHE II a severity of disease classification system. Crit Care Med 1985; 13: 818-29. 7. Barlow P, Murray L, Teasdale G. Outcome after severe head injury—the Glasgow model. In: Corbett WA, ed. Medical applications of microcomputers. New York: Wiley, 1987: 105-26.
Jennett WB. Inappropriate
1491
8. Stablein DM, Miller JD, Choi SC, Becker DP. Statistical methods for determining prognosis in severe head injury. Neurosurgery 1980; 6: 243-48. 9. Knaus WA, Rauss A, Alperovitch A, et al. Do objective estimates of chances for survival influence decisions to withhold or withdraw treatment? Med Decis Making 1990; 10: 163-71. 10. Jennett B, Bond M. Assessment of outcome after severe brain damage. Lancet 1975; i: 480-84. 11. Jennett B, Teasdale G, Braakman R, Minderhoud J, Knill-Jones R. Predicting outcome in individual patients after severe head injury. Lancet 1976; i: 1031-34. 12. Titterington DM, Murray GD, Murray LS, et al. Comparison of discrimination techniques applied to a complex data set of head injured patients. JR Statist Soc A 1981; 144: 145-75. 13. Spiegelhalter DJ. Evaluation of clinical decision aids with an application to dyspepsia. Statist Med 1983; 2: 207-16. 14. Murray GD, Murray LS, Barlow P, et al. Assessing the performance and clinical impact of a computerised prognostic system in severe head injury. Statist Med 1986; 5: 403-10. 15. Jennett B, Teasdale G, Fry J, et al. Treatment for severe head injury. J Neurol Neurosurg Psychiatry 1980; 43: 289-95. 16. Harari RJ, Narayan RK, Iacono L, Ishman R, Ghajar J. Marked
variability in the management of severe head injury at trauma centres in the United States. J Neurosurg 1992; 76: 397A. 17. Adams ID, Chan M, Clifford PC, et al. Computer aided diagnosis of acute abdominal pain: a multicentre study. BMJ 1986; 293: 800-04. 18. Wellwood J, Spiegelhalter DJ, Johannessen S. How does computer aided diagnosis improve the management of acute abdominal pain? Ann R Coll Surg Engl 1992; 74: 140-46. 19. Wyatt J. Lessons learned from the field trial of ACORN, an expert system to advise on chest pain. In: Barber B, Cao D, Qin D, eds. Proceedings of the sixth world conference on medical informatics, Singapore. Amsterdam: North Holland, 1989: 111-15. 20. Imbus SH, Zawacki BC. Autonomy for burned patients when survival is unprecedented. N Engl J Med 1977; 297: 308-10. 21. Cohen H. Response to automony for severely burned patients. N Engl J Med 1977; 297: 1182. 22. Gibson RM, Stephenson GC. Aggressive management of severe closed head trauma: time for reappraisal. Lancet 1989; ii: 369-71. 23. Feldman Z, Contant CF, Robertson CS, Narayan RK, Grossman RG. Evaluation of the Leeds prognostic score for severe head injury. Lancet 1991; 337: 1451-53. 24. de Dombal FT. Ethical considerations concerning computers in medicine in the 1980s. J Med Ethics 1987; 13: 179-84.
Influence of nutritional status on child rural Zaire
the association between nutritional is obvious for extreme malnutrition, the issue is not so clear for mild to moderate undernutrition. We have investigated this association in children of 0-5 years in the rural area of Bwamanda, Zaire, where an integrated development project, with good medical facilities, has operated for 20 years. A random cluster sample of 5167 children was taken; newborn infants and immigrants were included at six quarterly survey rounds from October, 1989, until February, 1991. All surveys included clinical and anthropometric assessment of nutritional status. Deaths were recorded up to April, 1992; there were 246 deaths. Marasmus, kwashiorkor, and other causes of death were defined by the verbal autopsy method and checked against medical records kept at the central hospital and the peripheral dispensaries. As expected, we found an increased risk of death in severe malnutrition. When deaths directly attributed to marasmus or kwashiorkor were excluded, mild to moderate stunting or wasting were not associated with higher mortality in the short term (within 3 months of the previous study round) or in the long term (from 3-30 months after study entry). The commonest causes of death were malaria and anaemia. Extreme marasmus and kwashiorkor caused 16% of deaths, and are important causes of death even in this favoured area with an integrated
Although
status and
mortality risk
development project. Nutritional interventions should be targeted more selectively so that children with moderate malnutrition can be protected from progression to marasmus or kwashiorkor.
mortality in
Introduction Our project is a semilongitudinal study among children aged 0-5 years of the relation between nutritional status and mortality. The existence of such a relation is obvious in extreme situations, such as disaster areas and refugee camps, where children clearly die from hunger. In most developing countries, a more appropriate question is whether mortality risk is also increased in mildly to moderately undernourished children who are not in hospital. If so, is it only because there is progression towards more severe malnutrition, or does mild to moderate malnutrition in itself increase the risk of death from non-nutritional (ie, mostly infectious) causes? The issue is still unresolved. Reports on the mortality risk associated with mild to moderate malnutrition are conflicting. There have been studies that did not show an increased mortality risk in moderate
malnutrition from Asia1 and from African Other studies have shown a gradual increase in risk of death with worsening nutrition.5-10 Our study was done in a relatively favoured area of Zaire, where an integrated development and health programme has been in operation for more than 20 years; the prevalence of diarrhoea and measles is low compared with many other parts of Africa. Our main aim was to assess, in a representative sample of young children, short-term and long-term mortality risks associated with different degrees of undernutrition, defined clinically as well as anthropometrically. The design differs from that of previous studies in that nutritional status at death was also taken into account. We decided to include this feature so that we could fmd out whether any increase in risk of death for mild malnutrition was due to worsening nutritional status. The study is part of a research project on maternal determinants of child health and care in a rural tropical area. ADDRESSES: Centre for Human Genetics (J. Van den Broeck, MD) and Department of Paediatrics (Prof R. Eeckels, MD), University of Leuven; and Institute of Tropical Medicine, Antwerp (Prof J. Vuylsteke, MD), Belgium. Correspondence to Dr J. Van den Broeck, Centre for Human Genetics, Herestraat 49, 3000 Leuven, Belgium.
12
June 1993
No 8859
ARTICLES
Does prediction of outcome alter patient
management?
A patient’s prognosis is a key factor for the clinicians involved in management. We set out to determine if provision of computer-based predictions of outcome after severe head injury resulted in measurable changes in patient management. In particular, we wondered whether introduction of the predictive system would alter the relation between severity of injury and "intensity" of
management.
patients admitted to four British neurosurgical units between 1986 and 1989 following a severe head injury, and who were either in coma for 6 h or had an operation for acute intracranial haematoma, were studied. Specified 1025
recorded and all patients were followed up after six months. The study had three phases: a baseline period of at least one year before the introduction of computerbased outcome prediction, one year when predictions were provided at specified times, and a final six months when prediction was withdrawn. While predictions were being provided, there was an increase in the use of specified aspects of intensive care in patients predicted to have a good outcome, but a 39% reduction in the use of these same aspects of intensive care in patients predicted to have the worst outcome. There was no evidence that the provision of predictions affected overall outcome, length of stay, or the recording of explicit decisions to limit treatment. We have demonstrated that the introduction of a routine prediction service can alter patient
aspects of intensive management
management.
were
Introduction The process by which a doctor arrives at a management decision for a seriously ill patient is often complex. Many factors have to be taken into consideration and balanced against one another. Information about the patient’s clinical state and items such as age and the results of biochemical
investigations
or
imaging procedures
are
important.
Resource availability is also relevant. For humanitarian, ethical, and economic reasons doctors are being encouraged to concentrate activities and resources on patients who are most likely to gain benefit, and to minimise intensive treatment for patients whose suffering is thereby prolonged or who would be left with an unacceptable quality of life. Conversely, undue pessimism, often based on a single criterion such as age, can lead to premature limitation of treatment, whereas unnecessary or unduly extended intensive care can expose patients with less severe injury to risks.l The expected eventual outcome is a crucial factor in making decisions about a patient’s management.2 Physicians’ estimates of outcome are often unduly optimistic or pessimistic or unnecessarily ambiguous,2-5 and a less subjective approach to the problem of predicting a patient’s future is required. Methods of predicting outcome in ADDRESSES: Department of Neurosurgery, University of Glasgow, Southern General Hospital, Glasgow G51 4TF, UK (L. S. Murray, PhD, Prof G. M. Teasdale, FRCS, Prof B. Jennett, MD, D. Kelly, BN); Department of Surgery, University of Glasgow, Western Infirmary, Glasgow (G. D. Murray, PhD); Department of Clinical Neurosciences, University of Edinburgh, Western General Hospital, Edinburgh (Prof J. D. Miller, MD, P. Jones, MAppSc); Wessex Neurological Centre, University of Southampton, Southampton General Hospital, Southampton (Prof J. D. Pickard, FRCS, S. Bailey, SRN); and The Walton Centre for Neurology and Neurosurgery NHS Trust, Walton Hospital, Liverpool (M. D. M. Shaw, FRCS, J. Achilles, RGN, J. Lacey, RGN). Correspondence to Dr L S. Murray.
1488
TABLE I-FEATURES OF PATIENTS IN DIFFERENT STUDY PERIODS ON ADMISSION TO NEUROSURGICAL UNITS
critically ill patients have been developed, but the effect on patient management of providing clinicians with this predictive information has not been widely examined. Knaus et al9 described changes in the pattern of treatment decisions that resulted from provision of information about the chances of survival of patients with multiple organ failure. Severe head injuries account for a substantial part of the work of neurosurgical units and many general intensive-care units. In many neurosurgical patients, early treatment efforts are rewarded by reasonable recovery, but even with the most intensive regimens of monitoring and therapy mortality for head-injured patients in coma is 30-40% and another 15-20% survive with persistent severe disability. We report the effect on patient management and outcome of providing clinicians with estimates of prognosis after severe head injury.
Patients and methods Patients The four centres that participated in the study-Glasgow, Edinburgh, Liverpool, and Southampton-provide a range of neurosurgical facilities for head-injured patients that embraces the
patterns of care in most UK regions. The patients studied were either in coma (ie, no eye opening, not speaking, and not obeying commands) for more than 6 h or had an acute intracranial haematoma that needed to be evacuated by craniotomy. Research nurses collected information on the clinical state of each patient throughout their stay, the investigations carried out, treatments given, the duration of treatments, and length of stay. Outcome was assessed at six months after injury according to the Glasgow outcome scale10 with three outcome groups: dead/ vegetative state, severe disability, and moderate disability/good recovery.
Assessment of patient management
Management was characterised by features that were deemed, on the basis of clinical judgment, to be those most likely to be influenced by a clinician’s perception of outcome. The use of invasive investigation (intracranial pressure monitoring), intensive treatment (intubation and ventilation, osmotic diuretics), and written instructions to limit treatment were noted. The time till death or, in survivors, the length of stay was also calculated. In addition to comparing the overall management practices, we postulated in advance that management might be influenced differently in cases with different severities of injury.
Experimental design The design adopted has been characterised as quasiexperimental." Thus, the study, which took place between 1986 and 1989, was divided into three periods: a baseline period of at least one year during which data were collected but no predictions of outcome given to staff, one year when the prediction program was used, and a final six months when prediction was withdrawn. Findings from the three periods were then compared. This approach was adopted because we believed that it was relevant to assess the overall performance of a neurosurgical department, and that randomisation of patients or clinicians or of the small number of centres would not be appropriate. The introduction of the prediction service was staggered at the four centres to take account of coincidental changes in practice due to other factors.
Before the system was introduced into each centre,
validity was established by comparing the predicted and actual outcome in patients who had been followed up. When the prediction service was introduced, the principles of calculating prognosis and the interpretation of probabilities were explained and the validity of the system demonstrated to staff at each centre. The observational nature of the study was emphasised and care was taken not to specify how predictions should be used to change practice. During the period when prediction was available, research nurses continued to collect patient data and also made predictions on the lst, 3rd, and 7th days after onset of coma, or before operation for those patients with an intracranial haematoma. Results of prediction were provided as a regular service, in a way similar to a laboratory report. Clinicians were made aware of results during clinical rounds. Medical and nursing staff had free access to the computer and were able to make predictions. If predictions were not made by ward staff at weekends they were provided by the research nurse as soon as possible afterwards. The prediction service was stopped after one year and the its
computer removed. Collection of data about new cases continued for a further six months and all cases were followed up for six months.
Statistical methods Outcome prediction The computer program that was used to calculate predicted outcome was designed so that it could be used by neurosurgical and nursing staff without special training, and run on a personal computer (Amstrad PCW8256). The program used an independence model based on Bayes’ theorem7,11 to update the probability of each outcome given the clinical information about the patient. First, the program requested patient and user identification and displayed appropriate prior probabilities of outcome. The user was then questioned about the patient’s age, findings of the Glasgow coma scale, changes in clinical condition, motor-response pattern, pupil reaction, and eye signs. After the entry of each item of clinical information, updated probabilities of outcome were calculated and displayed. Results were also presented by two graphical methods and printed copies provided. The decision to base predictions on the independence model was a pragmatic one, in that it was simple to make the predictions even when some prognostic variables were unavailable. This approach yielded predictions that were similar to those obtained by more sophisticated statistical modelling, the only disadvantage being that the predictions tended to be less well calibrated.’2
The effect on outcome of providing predictions was assessed by a logistic regression model that related the probability of a poor outcome (dead or vegetative) to the phase of study (before predictions vs during predictions), after correcting for the effects of age, pupil reaction, and Glasgow coma scale at admission to the neurological unit. Such a model has been shown previously, with similar
data,
to
give
an
excellent fit for the data, and
to
allow
development of a powerful, well calibrated scoring system for predicting outcome.14 The effect on patient management of providing outcome predictions was assessed first by dividing the patients into three prognostic groups, based on the logistic regression model, that were defined by the estimated probability of a poor outcome being 00 to 04 (good prognosis), 04 to 08 (moderate prognosis), or 08 to 10 (poor prognosis). The proportion of patients receiving various aspects of intensive management was then related to the prognostic group and to the phase of the study (before predictions vs during predictions) with a standard log-linear model. Particular emphasis was placed on the "interaction" term, which reflects any change in the relation between prognosis and management that was associated with the introduction of the prediction service.
1489
TABLE 11-OUTCOME SIX MONTHS AFTER INJURY IN DIFFERENT STUDY PERIODS
Percentages are shown unless otherwise
TABLE IV-USE OF INTENSIVE MONITORING AND TREATMENT IN DIFFERENT PROGNOSTIC GROUPS BY STUDY PERIOD
indicated.
These formal analyses did not incorporate the data from the final phase of the study, after the prediction service had been withdrawn. This was because the primary hypothesis of the study related to the effects of introducing the predictions, and, moreover, given the limited sample size for the final phase, any formal analysis of these
data would lack statistical power. However, data from the final
phase are reported to allow informal comparisons to be made. Results
patients were studied-38% in Glasgow, 25% in Edinburgh, 23% in Liverpool, and 14% in Southampton. The proportions of patients from the different centres and the types of patients admitted and their characteristics were similar in the different periods of the study (table l). In some patients it was not possible to make an outcome prediction because the use of paralysis and ventilation eliminated the physical signs necessary for the predictive system. This was the case in 1%, 19%, 16%, and 13% of patients pre-operatively, at 24 h, at 3 days, and at 7 days, respectively, during the period of the study. During the period when predictions were provided to medical staff, outcome prediction was available in 94% of appropriate 1025
cases.
6 months after injury in the different is shown in table II. Overall, just over study periods one-third of the patients died and almost half had a moderate disability or made a good recovery. The logistic regression model, which corrected for age and severity of injury, indicated no difference in outcome between the periods before and during which predictions were provided (p = 0-57, X2). This lack of effect of prediction on outcome was consistent over the three prognostic groups, with no substantial differences between the observed and expected numbers of poor outcomes in the three groups. There was an overall trend for a modest, but non significant, increase in intracranial-pressure monitoring, intubation and/or ventilation, and the administration of osmotic diuretics during the period when predictions were provided (table III). However, analysis of the interaction between initial severity and management showed variations in the different phases of the study. Thus, when predictions
Patient
outcome
TABLE III-USE OF ASPECTS OF MANAGEMENT IN DIFFERENT STUDY PERIODS I
I
I
*Patients are stratified ie, dead or vegetative.
according to their estimated probability of a
poor outcome-
provided, the pattern of use of intracranial-pressure monitoring, intubation and/or ventilation, and osmotic diuretics differed substantially compared with the period before prediction in patients in the same estimated prognosis group (table IV). Among patients whose features indicated a moderate or good prognosis, osmotics, intubation and/or ventilation, and intracranial-pressure monitoring were used more frequently during the period that predictions were provided. Conversely, during this period, there was a decrease of 39% in the average frequency of use of these aspects of intensive management in patients with the worst prognosis. This change in the relation between prognosis and management was statistically significant (X2 test) for the use of osmotics (p <0 01) and intubation and/or ventilation (p < 0-001). The pattern of redeployment of osmotics and intubation and/or ventilation away from patients with a poor prognosis and towards patients with a moderate or good prognosis was seen consistently across the four participating were
centres.
A decision to limit treatment was recorded in the case record of about 1 in 5 patients; this finding did not vary between the study periods, and was followed by a death or a dependent outcome in more than 9 out of 10 patients (95%, 98%, and 92% before, during, and after predictions, respectively). Also, there was no variation between the three periods in the time to death in patients with a fatal outcome or in the duration of stay in the neurosurgical unit of survivors (table III).
Discussion The development and evaluation of a mathematically based clinical decision aid can be likened to the phases in assessment of a new therapy. 13 In phase I the model is developed, in this case the independence model for predicting outcome in the early stage after severe head injury, based on easily obtained clinical signs.’1 In phase II, the quality of the information provided is assessed by comparing the results obtained from the predictive model with those of other statistical models and with eventual clinical outcome. We have reported previously that when clincial features are well selected, the independence model performed as well as, or even better than, more complex techniqueslz and gave predictions that were more consistent
1490
and reliable than those made by senior neurosurgeons.14 Our study represents the third phase of the development and evaluation process. It deals with the discovery of the practical consequences of providing staff in neurosurgical units with estimates of prognosis for severely head-injured patients. The system proved acceptable to staff, and its provision was associated with redeployment of some aspects of intensive care from patients with a poor prognosis to those with a better estimated prognosis. We assessed the effects of predictions by studying series of patients in different units in successive periods, rather than focusing on the treatment of individual patients or the actions of individual doctors, and we rejected the possibility of randomising patients for whom prognostic information would be provided. Neurosurgical units usually function with considerable cross-cover by junior doctors, particularly in emergency situations, so that a patient’s management is influenced by the unit’s general attitudes and policies. The mere presence of the predictive system in a unit could, therefore, have a confounding effect on the management of patients for whom predictions were not made. For similar reasons, the consequences of predictions on specific patients could not be separated from the general effect of the use of the system. The value of correction of hypotension and hypoxia, of the prompt evacuation of intracranial haematomas, and of the general intensive care of the unconscious head-injured patient is universally accepted. However, there is less agreement about the benefits of other components of intensive management of head injuries, and this is reflected in the variations in practice shown in this study. Thus, in different centres the proportion of patients treated with osmotic diuretics varied from 23% to 67%, with intubation and/or ventilation from 44% to 83%, and intracranialpressure monitoring from 1% to 57%. This finding is in accord with other reports15.16 of considerable variations in the use of these and other aspects of intensive management. When there are no clearly established indications for the use of a component of treatment, its employment may be governed by an explicit protocol or, more often, may be determined by specific decisions in each case. Our hypothesis that prognosis would be a factor in such decisions was confirmed by our findings. Thus, during the period when predictions were provided, osmotic diuretics, intubation and/or ventilation, and perhaps intracranial pressure monitoring were used less often in patients with a very poor prognosis whose outlook was unlikely to be altered, and more often in patients with either an uncertain or optimistic outlook who were more likely to benefit. This change was achieved partly by a slight overall increase in such treatments, but more so by redeployment with less use in cases with the worst prognosis. The absence of an alteration in outcome in association with these changes in treatment indicates that, as used in this study, these components of treatment did not have a crucial effect on outcome.
Our results, which indicate that clinicians respond appropriately to predictive information, should allay concern that an adverse prediction might result in unreasonable pessimism. The frequency with which explicit instructions were given to limit treatment varied from 13%
27% in the different units but was not related to the provision of predictions. Most fatalities occurred within 3 days of injury, and awareness of this early mortality may reduce the pressure on neurosurgical staff to take early decisions to limit treatment. to
Most clinical decision aids that have been subjected to phase III trial-the formal assessment of their influence on clinical practice-have been concerned with diagnosis and
the effects of their use have been variableY-19 However, Knaus et al9 studied patients with failure of three or more organ systems, and found that when objective estimates of a low chance of survival were provided, there was a small but significant increase in decisions to stop active treatment and to continue only with comfort care. The greatest effect was in the most seriously ill patients, in whom "comfort care only" decisions increased from 11 % to 36%, while the total proportion of patients receiving unlimited treatment remained stable. An earlier study20 of bum victims showed a similar effect of prediction, but was criticised as discouraging treatment before the patient’s response to treatment was known.21 Other systems for providing outcome predictions for individual head-injured patients have been devised in Richmond, Virginia, USA,8 and Leeds, UK,22 The Richmond system aims to provide information about the likelihood of different outcomes for a patient, but its influence on practice has not been reported. The Leeds score is based on data obtained within 12 h of admission to the neurosurgical unit, including the results of invasive monitoring, and it aims to identify patients whose death can be predicted with certainty. However, a study from Texas, USA, has shown that the Leeds model cannot be relied on "to clarify the complicated emotional, moral, legal, and financial issues that surround the early termination of care in
seriously head-injured patients" .23 We have examined the effect of providing clinicians with information about the expected outcome after severe head injury, but this approach could have a wider application in patient management in general. We believe that it is not realistic to expect a statistical predictive system to relieve clinical staff of their duty to take account of a range of factors when making decisions about the management of patients. These include concepts such as "error" and "harm" as well as clinical and laboratory data. An estimate of the likelihood of outcome is only one factor for the clinician to consider, and it has been likened to any other test.24 This study indicates that provision of predictive information can result in appropriate measurable changes in clinicial practice, without any evidence of adverse effects on outcome. We thank our neurosurgical colleagues for allowing us to study their patients and for their cooperation with the collection of data and provision of predictions. This study was supported by the Medical Research Council Special Project Grant G842699.
REFERENCES 1.
use of intensive care. BMJ 1984; 289: 1209-11. 2. Barlow P, Teasdale G. Predictions of outcome and the management of severe head injuries: the attitudes of neurosurgeons. Neurosurgery 1986; 19: 989-91. 3. Chang RWS, Lee B, Jacobs S. Accuracy of decisions to withdraw therapy in critically ill patients: clinical judgement versus a computer model. Crit Care Med 1989; 17: 1091-97. 4. Dawes RM, Faust D, Meehl RE. Clinical versus actuarial judgement. Science 1989; 243: 1668-74. 5. Kaufmann MA, Buchmann B, Scheidegger D, Gratzl O, Radu EW. Severe head injury: should expected outcome influence resuscitation and first-day decisions? Resuscitation 1992; 23: 199-206. 6. Knaus WA, Draper EA, Wagner DP, Zimmerman JE. APACHE II a severity of disease classification system. Crit Care Med 1985; 13: 818-29. 7. Barlow P, Murray L, Teasdale G. Outcome after severe head injury—the Glasgow model. In: Corbett WA, ed. Medical applications of microcomputers. New York: Wiley, 1987: 105-26.
Jennett WB. Inappropriate
1491
8. Stablein DM, Miller JD, Choi SC, Becker DP. Statistical methods for determining prognosis in severe head injury. Neurosurgery 1980; 6: 243-48. 9. Knaus WA, Rauss A, Alperovitch A, et al. Do objective estimates of chances for survival influence decisions to withhold or withdraw treatment? Med Decis Making 1990; 10: 163-71. 10. Jennett B, Bond M. Assessment of outcome after severe brain damage. Lancet 1975; i: 480-84. 11. Jennett B, Teasdale G, Braakman R, Minderhoud J, Knill-Jones R. Predicting outcome in individual patients after severe head injury. Lancet 1976; i: 1031-34. 12. Titterington DM, Murray GD, Murray LS, et al. Comparison of discrimination techniques applied to a complex data set of head injured patients. JR Statist Soc A 1981; 144: 145-75. 13. Spiegelhalter DJ. Evaluation of clinical decision aids with an application to dyspepsia. Statist Med 1983; 2: 207-16. 14. Murray GD, Murray LS, Barlow P, et al. Assessing the performance and clinical impact of a computerised prognostic system in severe head injury. Statist Med 1986; 5: 403-10. 15. Jennett B, Teasdale G, Fry J, et al. Treatment for severe head injury. J Neurol Neurosurg Psychiatry 1980; 43: 289-95. 16. Harari RJ, Narayan RK, Iacono L, Ishman R, Ghajar J. Marked
variability in the management of severe head injury at trauma centres in the United States. J Neurosurg 1992; 76: 397A. 17. Adams ID, Chan M, Clifford PC, et al. Computer aided diagnosis of acute abdominal pain: a multicentre study. BMJ 1986; 293: 800-04. 18. Wellwood J, Spiegelhalter DJ, Johannessen S. How does computer aided diagnosis improve the management of acute abdominal pain? Ann R Coll Surg Engl 1992; 74: 140-46. 19. Wyatt J. Lessons learned from the field trial of ACORN, an expert system to advise on chest pain. In: Barber B, Cao D, Qin D, eds. Proceedings of the sixth world conference on medical informatics, Singapore. Amsterdam: North Holland, 1989: 111-15. 20. Imbus SH, Zawacki BC. Autonomy for burned patients when survival is unprecedented. N Engl J Med 1977; 297: 308-10. 21. Cohen H. Response to automony for severely burned patients. N Engl J Med 1977; 297: 1182. 22. Gibson RM, Stephenson GC. Aggressive management of severe closed head trauma: time for reappraisal. Lancet 1989; ii: 369-71. 23. Feldman Z, Contant CF, Robertson CS, Narayan RK, Grossman RG. Evaluation of the Leeds prognostic score for severe head injury. Lancet 1991; 337: 1451-53. 24. de Dombal FT. Ethical considerations concerning computers in medicine in the 1980s. J Med Ethics 1987; 13: 179-84.
Influence of nutritional status on child rural Zaire
the association between nutritional is obvious for extreme malnutrition, the issue is not so clear for mild to moderate undernutrition. We have investigated this association in children of 0-5 years in the rural area of Bwamanda, Zaire, where an integrated development project, with good medical facilities, has operated for 20 years. A random cluster sample of 5167 children was taken; newborn infants and immigrants were included at six quarterly survey rounds from October, 1989, until February, 1991. All surveys included clinical and anthropometric assessment of nutritional status. Deaths were recorded up to April, 1992; there were 246 deaths. Marasmus, kwashiorkor, and other causes of death were defined by the verbal autopsy method and checked against medical records kept at the central hospital and the peripheral dispensaries. As expected, we found an increased risk of death in severe malnutrition. When deaths directly attributed to marasmus or kwashiorkor were excluded, mild to moderate stunting or wasting were not associated with higher mortality in the short term (within 3 months of the previous study round) or in the long term (from 3-30 months after study entry). The commonest causes of death were malaria and anaemia. Extreme marasmus and kwashiorkor caused 16% of deaths, and are important causes of death even in this favoured area with an integrated
Although
status and
mortality risk
development project. Nutritional interventions should be targeted more selectively so that children with moderate malnutrition can be protected from progression to marasmus or kwashiorkor.
mortality in
Introduction Our project is a semilongitudinal study among children aged 0-5 years of the relation between nutritional status and mortality. The existence of such a relation is obvious in extreme situations, such as disaster areas and refugee camps, where children clearly die from hunger. In most developing countries, a more appropriate question is whether mortality risk is also increased in mildly to moderately undernourished children who are not in hospital. If so, is it only because there is progression towards more severe malnutrition, or does mild to moderate malnutrition in itself increase the risk of death from non-nutritional (ie, mostly infectious) causes? The issue is still unresolved. Reports on the mortality risk associated with mild to moderate malnutrition are conflicting. There have been studies that did not show an increased mortality risk in moderate
malnutrition from Asia1 and from African Other studies have shown a gradual increase in risk of death with worsening nutrition.5-10 Our study was done in a relatively favoured area of Zaire, where an integrated development and health programme has been in operation for more than 20 years; the prevalence of diarrhoea and measles is low compared with many other parts of Africa. Our main aim was to assess, in a representative sample of young children, short-term and long-term mortality risks associated with different degrees of undernutrition, defined clinically as well as anthropometrically. The design differs from that of previous studies in that nutritional status at death was also taken into account. We decided to include this feature so that we could fmd out whether any increase in risk of death for mild malnutrition was due to worsening nutritional status. The study is part of a research project on maternal determinants of child health and care in a rural tropical area. ADDRESSES: Centre for Human Genetics (J. Van den Broeck, MD) and Department of Paediatrics (Prof R. Eeckels, MD), University of Leuven; and Institute of Tropical Medicine, Antwerp (Prof J. Vuylsteke, MD), Belgium. Correspondence to Dr J. Van den Broeck, Centre for Human Genetics, Herestraat 49, 3000 Leuven, Belgium.