Received 25th August 2004; returned for revisions 14th December 2004; revised manuscript accepted 3rd January 2005.
(e.g., six months after stroke). Stroke patients To explore more simple modelling methods, we
typically, however, show nonlinear recovery pat- investigated mathematical powers, logarithms,
terns.14,15 In most stroke cases, patients show rapid double-logarithms, and other simple mathematical
recovery during the initial few months, after which functions. Of those, we focused on natural loga-
the pace of recovery to six months from onset rithmic functions (ln) because they displayed three
slows towards the final outcome.16 Consequently, advantages for modelling functional recovery.
linear modelling is not up to the task of accurately First, the progress curves (Figure 1) resembled
predicting the prospective outcome. To simulate actual recovery patterns: if the recovery target is
the nonlinear aspects of functional recovery, neu- set at 180 days from onset and assigned a value of
ral network modelling,17 logistic modelling,18,19 100%, approximately 70% of recovery is registered
and other types of nonlinear modelling have been at 90 days and subsequent progress occurs at a
proposed. Although more successfully predictive, reduced rate.1,16,20 Similarly, a logarithmic func-
these modelling methods are not widely applied tion fitted the recovery patterns of upper limb
because of their mathematical complexity. Thus, function of stroke patients.21 Second, based on
for general clinical applicability, there has been a scores sampled on two days separated by an
need for a simpler means of accurately predicting interval, using simple mathematical procedures
the progress of recovery. (Figure 1), the modelling formula can easily be
Figure 1 Model formula and predictive curve. (A) shows a generic structure; (B) shows mathematical procedures to tailor the
generic structure to fit individual degree of recovery. For this, actual FIM scores recorded at two time points (Day A and Day B)
are required. DFIM indicates change in FIM scores between Day A and Day B. Constant in (A) is countervailed in this procedure.
(C) shows the final form of the model formula. Predicted value for Day X can be calculated with this form. FIM, Functional
Independence Measure; ln, natural logarithm.
Logarithmic modelling in hemiplegic stroke 781
scaled to fit each individual’s magnitude of recov- involve motor functions (FIM-motor) and other
ery. Third, owing to mathematical specificity of two concern cognitive functions (FIM-cognition).
logarithms, the model formula can easily be The total scores score for all 18 items (FIM-
calculated (e.g., ln(90)/ln(30)/ln(90/30) /ln(3), total) is commonly used to assess functional
see Figure 1B). To evaluate the practical usefulness independence in rehabilitation medicine (totally
of logarithmic modelling we carried out a long- dependent in ADL /18, completely independent
itudinal study. in ADL /126).
Using FIM scores, nursing staff assessed the
functional recovery of patients in terms of ADL.
Methods Evaluations were typically recorded a few days
after admission, again at two to six weeks after
Patients admission, and then once a month during hospi-
Stroke patients with hemiplegia who were ad- talization. In our study, to assure reliability of the
mitted to our long-term rehabilitation hospital evaluations, FIM scores were reviewed at weekly
during August 2003 to April 2004 were recruited conferences.
into the study. Criteria for inclusion were:
no past history of hemiplegia; capable of indepen-
dent ADL (activities of daily life) before stroke; Modelling and evaluation
wheelchair required for locomotion at admission. A generic structure of modelling was given in a
As a result of Japanese health insurance proce- simple natural logarithmic formula (independent
dures, patients were referred from local community variable /days from onset) (Figure 1A). To tailor
acute medical services, typically 30/60 days after the generic structure to fit each individual’s degree
the stroke occurred, and received inpatient care in of functional recovery, we performed calculations
our long-term rehabilitation hospital for 30 /180 on the total FIM scores at the first two time-points
days. During the prior period of acute medical
after admission. For each patient, the increase in
hospitalization they received physical therapy.
total FIM scores between these two time-points
During long-term rehabilitation hospitalization
(DFIM) was used as the basis for scaling a co-
they received physical therapy, occupational ther-
apy and speech therapy for a joint total of 120 min efficient (b) in the generic structure (Figure 1B).
every day. To minimize the influence of variability The introduction of this countervailed the constant
of therapeutic regimen, we also limited recruitment in the generic structure. Thus, using the scores at
to patients who received treatment from the same the initial two sampling points, a generic struc-
rehabilitation team directed by a single physiatrist ture could be tailored to forecast each patient’s
(first author of this article). The protocol was functional recovery (model formula shown in
reviewed and approved by our hospital’s ethical Figure 1C).
committee and informed consent was obtained To assess the fit of the time-course of the
from all patients. model, FIM scores (FIM-total, FIM-motor and
FIM-cognition) were, on an individual basis,
Assessment of functional recovery longitudinally plotted with predicted values for
To assess functional recovery, we employed the each patient derived from the model formula. To
Functional Independence Measure (FIM), which assess the general applicability of logarithmic
has been widely used in rehabilitation medicine.22 modelling, using data from all patients, a conven-
The FIM is derived from scoring 18 items accord- tional linear regression analysis was performed to
ing to a seven-point scale (1 /totally dependent, compare the total FIM scores that were actually
7/completely independent) to assess functional obtained and the predicted values that were
independence in ADL. These 18 items are derived from the model formula. For this analysis,
categorized as self-care (6 items), sphincter we excluded the scores obtained at the first
control (2 items), transfers (3 items), locomotion two sampling points (indicated by arrowheads
(2 items), communication (2 items), and social in Figures 2 and 3) to determine the particular b
cognition (3 items). The first four categories coefficients.
782 T Koyama et al.
Figure 2 Time-course of actually obtained and predicted FIM scores for patients with left hemisphere lesions (cases 1 /10).
Closed circles show actually obtained FIM-total scores, closed triangles show actually obtained FIM-motor scores, closed
squares show actually obtained FIM-cognition scores and open circles show predictive values derived from the model formula
(Figure 1). Arrowheads indicate initial two sampling time-points for data to tailor the model formula for each individual.
FIM, Functional Independence Measure.
Logarithmic modelling in hemiplegic stroke 783
Figure 3 Time-course of actually obtained and predicted FIM scores for patients with right hemisphere lesions (cases 11 /18).
Closed circles show actually obtained FIM-total scores, closed triangles show actually obtained FIM-motor scores, closed
squares show actually obtained FIM-cognition scores and open circles show predictive values derived from the model formula
(Figure 1). Arrowheads indicate initial two sampling time-points for data to tailor the model formula for each individual.
FIM, Functional Independence Measure.
Case Age Gender Hemisphere Lesion Cause of stroke Ope. Intervention Comorbidity
CAD, coronary artery disease; DM, diabetes mellitus; HCC, hepatic cell carcinoma (post operation); HL, hyperlipidaemia; HT,
hypertension; MCA, middle cerebral artery; Ope., operation (open-skull) during acute medical hospitalization; OT, occupational
therapy; PT, physical therapy; ST, speech therapy.
(median, 63.5), motor FIM scores ranged from 14 of FIM scores were sampled at from 46 to 104 days
to 74 (median, 36), and cognition FIM scores (median 72) after occurrence (indicated by arrow-
ranged from 6 to 35 (median, 25). Initial FIM heads in Figures 2 and 3). The interval between
scores were sampled at from 32 to 77 days (median these two time-points ranged from 13 to 44 days
50) after occurrence of stroke and the second set (median 32).
Figure 4 Scatterplots showing the relationships between actually obtained FIM-total scores and predicted values derived
from the model formula (see Figures 2 and 3). Data from the two initial sampling time-points for each patient (indicated by
arrowheads in Figures 2 and 3) were excluded from the scatterplots. FIM, Functional Independence Measure.
Logarithmic modelling in hemiplegic stroke 785
model accurately predicted the actual results later The model formula is simple and structurally
obtained for the individuals. We know of no other flexible (Figure 1C). For consistency, in this study,
prediction modelling studies that provide useful, we used data from the first and second FIM
simple, individual-based mathematical modelling. samplings after admission. Any pair of periodic
When forecasting the functional recovery of an samplings, however, are suitable for defining the
individual stroke patient, a single physiatrist often coefficient (b) of the model formula. The flexibility
takes many clinical parameters into consideration. of the model formula enables easy re-estimation if
These include: initial motor and cognitive impair- predictive and actual values deviate. This simpli-
ment levels,26,27 initial day of rehabilitation,28 city and flexibility means that the model formula is
recovery rate,29 site and size of lesion,30,31 age,32 suitable for wide clinical application.
psychological status,33 unilateral spatial neglect,34
co-morbidities35 and other factors.36 Most of these Possible limitations of logarithmic modelling
previous prediction studies have attempted to In this study, we customized the individual’s
integrate multiple factors into the model. Our model formula by using scores from two FIM
study, however, uses only FIM scores sampled on samples: based on results of assessment done with
different days with an interval of 2/6 weeks an intervening period of from 13 to 44 days, data
between them. The results show that, processed from this sampling pair were collected at between
through our logarithmic equation, these data 32 and 104 days after the occurrence of stroke. The
enable powerful and accurate forecasting of func- model was effective within these sampling para-
tional recovery. Since initial patterns of recovery meters. Further studies are needed to find out the
could be affected by any of the multiple factors limits of applicability to FIM data collected at
mentioned above, the FIM scores of individual earlier or later phases of affliction.37 It is promising
patients are likely to be influenced by some or all that case 15 (Figure 3), using data collected
of these factors. relatively soon (33 days) after stroke occurrence
and with a short sampling interval (13 days),
provided accurate prediction. This modelling
Simplicity of logarithmic modelling might be useful even at earlier stages of illness
With the goal of developing a new forecasting and during shorter periods of hospitalization.
technique to predict functional recovery, we tested Close observation of the time-course data
several mathematical functions in our preliminary plotted for each individual revealed that factors
analyses. Taking a lead from a previous study that for change in the predictive model were the motor
employed logarithmic transformations of FIM- components rather than cognitive components.
total scores to model functional outcome,12 we Thus the model may not be applicable for patients
tested, among other manipulations, various dou- whose clinical manifestations are mainly cognitive
ble-logarithmic functions. In fact, in some cases, rather than motor (e.g., patients with subara-
preliminary models using double-logarithmic func- chnoid haemorrhage).38,39 Time-course plotting
tions did fit actual data slightly better than the also revealed a tendency for predicted values to
model formula that we are presenting here. Even exceed the actual data towards the high end of the
so, we preferred not to employ double-logarithmic FIM-total range. In view of the linearity of raw
modelling because of its complexity. FIM-scores (as discussed above) this might imply
Focusing on logarithmic modelling, we at- that the model formula is best utilized when
tempted to adjust the clause for ‘days from onset’ predictive values range from 25 to 120.
to improve the model fit. We attempted adjustment
based on the clinical observation that the start of Applicability of logarithmic modelling
functional recovery varies from case to case Our study samples yielded data on patients who
depending on site, size, and age of lesion. Our varied widely in age, lesion characteristics, and
preliminary analyses, however, revealed that the levels of motor and cognitive disabilities. The
contribution of such adjustments was minimal. results that we obtained indicate that the logarith-
Thus, to keep things simple, we applied logarithmic mic model formula (Figure 1C) could be effectively
modelling without any adjustments (Figure 1). applied for various types of hemiplegic stroke
Logarithmic modelling in hemiplegic stroke 787
patients. Our new model is valuable for its independence in stroke patients admitted to a
simplicity and applicability on an individual basis. rehabilitation programme. Clin Rehabil 1999; 13:
Using FIM scores that were sampled at two 464 /75.
different time-points, using a regular pocket calcu- 11 Thommessen B, Bautz-Holter E, Laake K.
Predictors of outcome of rehabilitation of elderly
lator (without a log function) and a logarithm
stroke patients in a geriatric ward. Clin Rehabil
look-up table (see Appendix), within minutes it is 1999; 13: 123 /28.
possible to come up with a prediction for each 12 Inouye M. Predicting outcomes of patients in Japan
individual’s functional status for a particular day. after first acute stroke using a simple model. Am J
Thus, the model formula, based on simple loga- Phys Med Rehabil 2001; 80: 645 /49.
rithmic function, could be adopted in everyday 13 Giaquinto S, Buzzelli S, Di Francesco L et al . On
clinical practice for predicting the functional the prognosis of outcome after stroke. Acta Neurol
recovery of stroke patients with hemiplegia. Scand 1999; 100: 202 /208.
14 Partridge CJ, Johnston M, Edwards S. Recovery
from physical disability after stroke: normal
patterns as a basis for evaluation. Lancet 1987; 1:
References 373 /75.
15 Jorgensen HS, Nakayama H, Raaschou HO,
Vive-Larsen J, Stoier M, Olsen TS. Outcome and
1 Tilling K, Sterne JA, Rudd AG, Glass TA, Wityk
time course of recovery in stroke. Part II: Time
RJ, Wolfe CD. A new method for predicting
course of recovery. The Copenhagen Stroke Study.
recovery after stroke. Stroke 2001; 32: 2867 /73.
2 Dam M, Tonin P, Casson S et al . The effects of Arch Phys Med Rehabil 1995; 76: 406 /12.
long-term rehabilitation therapy on poststroke 16 Sonoda S, Chino N, Domen K, Saitoh E. Changes
hemiplegic patients. Stroke 1993; 24: 1186 /91. in impairment and disability from the third to the
3 Allen CM. Predicting the outcome of acute stroke: a sixth month after stroke and its relationship
prognostic score. J Neurol Neurosurg Psychiatry evaluated by an artificial neural network. Am J Phys
1984; 47: 475 /80. Med Rehabil 1997; 76: 395 /400.
4 Reding MJ, Potes E. Rehabilitation outcome 17 Oczkowski WJ, Barreca S. Neural network
following initial unilateral hemispheric stroke. Life modelling accurately predicts the functional
table analysis approach. Stroke 1988; 19: 1354 /58. outcome of stroke survivors with moderate
5 Gladman JR, Harwood DM, Barer DH. Predicting disabilities. Arch Phys Med Rehabil 1997; 78:
the outcome of acute stroke: prospective evaluation 340 /45.
of five multivariate models and comparison with 18 Stineman MG, Maislin G, Fiedler RC, Granger CV.
simple methods. J Neurol Neurosurg Psychiatry A prediction model for functional recovery in
1992; 55: 347 /51. stroke. Stroke 1997; 28: 550 /56.
6 Falconer JA, Naughton BJ, Dunlop DD, Roth EJ, 19 Lofgren B, Gustafson Y, Nyberg L.
Strasser DC, Sinacore JM. Predicting stroke Cross-validation of a model predicting discharge
inpatient rehabilitation outcome using a home after stroke rehabilitation. Validating stroke
classification tree approach. Arch Phys Med Rehabil discharge predictors. Cerebrovasc Dis 2000; 10:
1994; 75: 619 /25. 118 /25.
7 Gompertz P, Pound P, Ebrahim S. Predicting stroke 20 Calautti C, Baron JC. Functional neuroimaging
outcome: Guy’s prognostic score in practice. studies of motor recovery after stroke in adults: a
J Neurol Neurosurg Psychiatry 1994; 57: 932 /35. review. Stroke 2003; 34: 1553 /66 Epub 2003 May 8.
8 Mauthe RW, Haaf DC, Hayn P, Krall JM. 21 Goodwin N, Sunderland A. Intensive, time-series
Predicting discharge destination of stroke patients measurement of upper limb recovery in the
using a mathematical model based on six items from subacute phase following stroke. Clin Rehabil 2003;
the Functional Independence Measure. Arch Phys 17: 69 /82.
Med Rehabil 1996; 77: 10 /13. 22 Linacre JM, Heinemann AW, Wright BD, Granger
9 Lai SM, Duncan PW, Keighley J. Prediction of CV, Hamilton BB. The structure and stability of the
functional outcome after stroke: comparison of the Functional Independence Measure. Arch Phys Med
Orpington Prognostic Scale and the NIH Stroke Rehabil 1994; 75: 127 /32.
Scale. Stroke 1998; 29: 1838 /42. 23 Wright BD, Linacre JM, Smith RM, Heinemann
10 Sanchez-Blanco I, Ochoa-Sangrador C, AW, Granger CV. FIM measurement properties and
Lopez-Munain L, Izquierdo-Sanchez M, Rasch model details. Scand J Rehabil Med 1997; 29:
Fermoso-Garcia J. Predictive model of functional 267 /72.
788 T Koyama et al.
24 Ring H, Feder M, Schwartz J, Samuels G. in first stroke patients. Arch Phys Med Rehabil 1994;
Functional measures of first-stroke rehabilitation 75: 858 /60.
inpatients: usefulness of the Functional 32 Bagg S, Pombo AP, Hopman W. Effect of age on
Independence Measure total score with a clinical functional outcomes after stroke rehabilitation.
rationale. Arch Phys Med Rehabil 1997; 78: 630 /35. Stroke 2002; 33: 179 /85.
25 Kwon S, Hartzema AG, Duncan PW, Min-Lai S. 33 Paolucci S, Antonucci G, Pratesi L, Traballesi M,
Disability measures in stroke: relationship among Grasso MG, Lubich S. Poststroke depression and
the Barthel Index, the Functional Independence its role in rehabilitation of inpatients. Arch Phys
Measure, and the Modified Rankin Scale. Stroke Med Rehabil 1999; 80: 985 /90.
2004; 35: 918 /23. 34 Katz N, Hartman-Maeir A, Ring H, Soroker N.
26 Ween JE, Alexander MP, D’Esposito M, Roberts Functional disability and rehabilitation outcome in
M. Factors predictive of stroke outcome in a right hemisphere damaged patients with and
rehabilitation setting. Neurology 1996; 47: 388 /92. without unilateral spatial neglect. Arch Phys Med
27 Zinn S, Dudley TK, Bosworth HB, Hoenig HM, Rehabil 1999; 80: 379 /84.
Duncan PW, Horner RD. The effect of poststroke 35 Pettersen R, Dahl T, Wyller TB. Prediction
cognitive impairment on rehabilitation process and of long-term functional outcome after stroke
functional outcome. Arch Phys Med Rehabil 2004; rehabilitation. Clin Rehabil 2002; 16: 149 /59.
85: 1084 /90. 36 Kwakkel G, Wagenaar RC, Kollen BJ, Lankhorst
28 Novack TA, Satterfield WT, Lyons K, Kolski G, GJ. Predicting disability in stroke /a critical review
Hackmeyer L, Connor M. Stroke onset and of the literature. Age Ageing 1996; 25: 479 /89.
rehabilitation: time lag as a factor in treatment 37 Asberg KH, Nydevik I. Early prognosis of stroke
outcome. Arch Phys Med Rehabil 1984; 65: 316 /19. outcome by means of Katz Index of activities of
29 Mayo NE, Korner-Bitensky NA, Becker R. daily living. Scand J Rehabil Med 1991; 23: 187 /91.
Recovery time of independent function post-stroke. 38 Hellawell DJ, Taylor R, Pentland B. Persisting
Am J Phys Med Rehabil 1991; 70: 5 /12. symptoms and carers’ views of outcome after
30 Chaudhuri G, Harvey RF, Sulton LD, Lambert subarachnoid haemorrhage. Clin Rehabil 1999; 13:
RW. Computerized tomography head scans as 333 /40.
predictors of functional outcome of stroke patients. 39 Svensson E, Starmark JE. Evaluation of individual
Arch Phys Med Rehabil 1988; 69: 496 /98. and group changes in social outcome after
31 Saeki S, Ogata H, Hachisuka K, Okubo T, aneurysmal subarachnoid haemorrhage: a long-
Takahashi K, Hoshuyama T. Association between term follow-up study. J Rehabil Med 2002; 34:
location of the lesion and discharge status of ADL 251 /59.
Logarithmic modelling in hemiplegic stroke 789
x Ln (x )
1.0 0.000
1.1 0.095
1.2 0.182
1.3 0.262
1.4 0.336
1.5 0.405
1.6 0.470
1.7 0.531
1.8 0.588
1.9 0.642
2.0 0.693
2.2 0.788
2.4 0.875
2.6 0.956
2.8 1.030
3.0 1.099
4.0 1.386
5.0 1.609
6.0 1.792
7.0 1.946
8.0 2.079
9.0 2.197
10.0 2.303