- Email: [email protected]
Analysis of neurological sequelae from radiosurgery of arteriovenous malformations: How location affects outcome
PI1 SO360-3016(97)00718-9
ELSEVIER
l
Clinical
Investigation
ANALYSIS OF NEUROLOGICAL SEQUELAE FROM RADIOSURGERY OF ARTERIOVENOUS MALFORMATIONS: HOW LOCATION AFFECTS OUTCOME JOHN
C. FLICKINGER, M.D.,*’
DOUGLAS KONDZIOLKA, M.D.,*” L. DADE LUNSFORD, M.D. *-c
ANN H. MART,
M.S,‘.
‘ANI?
Departments of *Radiation Oncology.+ Neurological Surgery and *Radiology. University of Pittsburgh School of Medicine. Pittsburgh, PA Purpose/Objective: To elucidate how the risks of developing temporary and permanent neuroIogicaI sequelae tram radIosurgery for arteriovenous maIformaHons (AVM) are related to AVM location, the addition of stereotactic magnetic resonance (MR) imaging to angiographic targeting, and prior hemorrhage or neurdogical deficits. Ma&h& and Methods: We evaluated follow-up imaging and clinical data in 332 AVM patients who received gmIIU3 knife radiosurgery at the University of Pittsburgh between 1987 and 1994. All patients had regular clinical or imaging follow-up for a minimum of 2 years (range: 24-96 months, median = 45 months}, There were 83 patients with MR-assisted planning, 187 with prior hemorrhages, and 143 with prior neurological deficits. ResuIts: Symptomatic postradlosurgery sequelae (any neurological problem including headache) developed in 30 (996)f 332 patients. Symptoms resdved in 58% of patients within 27 months with a sigu&a&Iy greater proportion @ = 0.006) resolving in patients with Dmin < 20 Gy vs. 2 20 Gy (89 vs. 36%). The 7-year actuarial rate for developing persistent symptomatic sequelae was 3.8 % . We first evaluated the reIatIve risks for -rent locations to construct a postradiosurgery injury expression (PIE) score for AVM location. Multivariate Iogistic regression analysis of symptomatic postradlosurgery sequeiae identified independent signiIicant correlations with PIE location score (p = O.@tMY7) and 12 Gy volume (p = 0.008), but with none of the other factors tested (p > 0.3). including the add&m of MR targeting, average radiation dose in 20 cc, prior hemorrhage, or ueurok@d deficit. We used these results to construct a risk prediction model for symptomatic postradiosurgery sequelsre. The risk of radiation necrosis was s@i&antIy correlated with PIE score (p < O&48), but not with 12-Gy v&tune. Conclusion: The risks of developing complications from AVM radiosurgery can be predicted according to location with the PIE score, in conjunction with the 12-Gy treatment volume. Further study of factors affecting persistence of these sequelae (progression to radiation necrosis) is needed. 0 1998 Elsevier Science Inc. Radiosurgery, Stereotactic surgery, Arteriovenous
malformation,
Complications, Radiation injury.
studies of injury responses after radiosurgery. In our first AVM study, we were able to significantly correlate postradiosurgery imaging changes (with and without symptoms) only to treatment volume, but not the to integrated logistic formula predictions (6). A second finding was that brainstem location was significantly associated with the development of symptoms in patients developing imaging changes (6). In a second study that also included benign tumors, we documented higher rates of postradiosurgery imaging (PRI) changes with AVM !31%1) than with a meningioma or acoustic tumors (8%), indicating the importance of tissue response from within the radiosurgery target (hemodynamic or otherwise) in contributing to the injury response. Two new tolerance models that described the tolerance of inhomogeneous radiosurgery dose distributions by the volume of a single minimum threshold dose were proposed in this study (3, 7). The logistic threshold-dose vol-
INTRODUCTION Radiosurgery has proven to be a dramatically successful technique for treating arteriovenous malformations by achieving overall high obliteration rates with relatively low morbidity. Dose-selection for radiosurgery mandates balancing the competing risks of higher obliteration rates (and lower risks of death or morbidity from future hemorrhage) for high doses, against lower risks of morbidity from radiation injury with lower doses. The 1% dose-volume isoeffeet line of Kjellberg et al. and the 3% dose-volume isoeffeet curve from the integrated logistic formula for predicting brain radiation necrosis are widely used radiosurgery doseprescription tolerance guidelines that were based on limited data available at the time (2, 10, 12). We now have more clinical data that can be used to refine tolerance predictions. Several important lessons were learned from our prior Reprint requests to: John C. Flickinger, M.D., Joint Radiation Oncology Center, 200 Lothrop Street. Pittsburgh, PA 1.5213.
Accepted for publication 3 I July 1997 273
274
I. J. Radiation Oncology l Biology l Physics
ume model (which used the volume enclosed within a minimum threshold-dose) provided the closest fit to the data (3, 7). Our most recent analysis of imaging changes after AVM radiosurgery revealed several new findings (5). We found that the total volume of tissue receiving 12 Gy or more (the 12-Gy volume or V12) was the only independently significant factor in multivariate logistic regression analysis that could be used to accurately predict the risks of developing all postradiosurgical imaging changes. The resulting model was, therefore, a logistic threshold-dose model with a threshold dose of 12 Gy. Voges et al. used a similar approach to correlate postradiosurgery imaging changes in 56 AVM and 77 tumor patients to a threshold dose of 10 Gy (16). In addition to the correlation of all PRI changes with 12-Gy volume, our study found that symptomatic postradiosurgery imaging changes were correlated with both 12-Gy volume and location, simply classified as either brainstem or nonbrainstem (5). We also found that symptomatic postradiosurgery changes were less likely to resolve than asymptomatic PRI changes (53% vs. 95%). Because of this difference, we concluded that future studies should focus specifically on symptomatic postradiosurgery changes. Several important questions arose, shortly after we completed the last study of AVM radiosurgery complications, that we investigated in this new analysis: 1. Did the addition of stereotactic MRI to angiographic imaging for radiosurgery treatment planning reduce the risk of complications? 2. Did patients with prior neurological deficits and/or prior hemorrhage have a lower risk of symptomatic postradiosurgery sequelae? 3. How much does the risk of symptomatic postradiosurgery injury or radiation necrosis vary with brain location? 4. Does the maximum average dose of 20 cc as described by Lax and Karlsson (11) improve upon or complement logistic risk prediction of postradiosurgery injury with 12-Gy volume? 5. Can the risks for permanent symptomatic postradiosurgery injury (radiation necrosis) be predicted by simply multiplying prediction model risks for all symptomatic postradiosurgery sequelae by the overall chance of symptom persistence, or is there some difference in the dose-volume relationship between temporary and permanent injuries? MATERIALS
AND METHODS
Clinical material and treatment parameters
We evaluated follow-up MRI scans and clinical data in 332 patients who received gamma knife radiosurgery at the University of Pittsburgh between August 1987 and August 1994. Approval for this study was granted by the University of Pittsburgh Institutional Review Board. All patients had
Volume 40, Number 2, 1998
Table 1. Treatment parameters in 332 AVM patients Standard Variable Dmin (GY) Qn,
(GY)
Isodose (%) Isocenters (#) Volume (cc) 12-Gy volume (cc)
Mean
deviation
Range
20.9 37.2 57.4 2.23 4.38 8.47
3.5 7.3 11.1 1.62 3.83 5.97
12-30 22-59 35-90 l-11 0.01-26.3 0.64-41.6
regular clinical or imaging follow-up for a minimum of 2 years (median = 45 months, range: 24-96 months) at the time of data analysis, Thirty-five patients had no MRI exams between 1 and 2 years after radiosurgery, but did have adequate clinical follow-up. They were included for the analysis of symptomatic postradiosurgery sequelae, but were excluded from the analysis of all PRI changes (this left 297 patients with adequate imaging follow-up for analysis). Prior to radiosurgery, AVMs bled in 187 patients. We documented preexisting neurological deficits (not including histories of headache or seizures) the day before radiosurgery in 143 patients from either AVM hemorrhage (n = 116) or prior therapeutic intervention (n = 27). A total of 220 of 332 patients had an AVM-related neurological deficit and/or a history of AVM hemorrhage. Treatment parameters are listed in Table 1. We planned radiosurgery for 83 patients using both stereotactic MR and angiographic imaging. In the other 249 patients, the only stereotactic imaging performed for treatment planning was biplane angiography. We referred to prior nonstereotactic computed tomographic and MR scans for supplemental 3D information to aid shaping of the target volume from the stereotactic biplane angiography. Dose-volume histograms were reviewed to calculate 12-Gy volumes (V12), which we defined as the total volume of all tissue (including the AVM target) receiving an equal or greater radiation dose than 12 Gy. Cumulative dose-volume histograms were also used to calculate the maximum average dose (D,,,) within 20 cc in the tissue around and including the target (11). Statistical analysis
The primary endpoint for statistical analysis was the development of symptomatic postradiosurgery sequelae. This was defined as any new neurological symptoms, including headache, that developed after radiosurgery that could not be attributed to hemorrhage. We also studied the development of all postradiosurgery imaging (PRI) changes as a secondary endpoint. This was defined as the development of new regions of long-TR imaging changes on postradiosurgery follow-up MR imaging, whether or not these were accompanied by symptoms. The final endpoint examined was persistent symptomatic postradiosurgery sequelae or symptomatic brain necrosis. This was defined as the documented persistence of symptomatic postradiosurgery sequelae for more than 2 years after the onset of symptoms. A p value of 0.05 was considered significant for all
AVM radiosurgery sequelae l
statistical tests. Stepwise multivariate logistic regression analysis of the effects of treatment variables on outcome was performed using BMDP software (1). Only variables approaching statistical significance (p < 0.10) were allowed in the final regression models for symptomatic postradiosurgery sequelae and for all PRI changes. We modeled the endpoint of radiation necrosis using a previously described actuarial correction averaging method for logistic regression analysis (4). Actuarial rates for the persistence long-TR imaging changes were calculated using the method of Kaplan and Meier (9). Statistical comparison between actuarial curves was performed with the log-rank test (14). RESULTS Incidence of postradiosurgical sequelae
Symptomatic postradiosurgery sequelae (any neurological problem including headache) developed in 30 (9.0%) of 332 patients. Symptoms resolved in 17 of 30 patients from 4 to 27 months after onset. Symptoms persisted for more than 24 months in 6 of 30 patients (1.8% of all patients), and were still present in 7 of 30 patients with less than 24 months of follow-up after symptom onset. The actuarial rate of symptom resolution was 57.6 ? 11.3% beyond 26 months after onset. The 7-year actuarial rate for developing persistent symptomatic sequelae was 3.8% (9.0% symptomatic X 42.4% unresolved). We identified PRI changes in 90 (30.3%) of 297 patients with regular imaging follow-up. Asymptomatic postradiosurgery imaging changes developed in 21.3% of all patients who had radiosurgery. Multivariate analysis and modeling
We first evaluated the relative risks of symptomatic postradiosurgery sequelae for different locations within the brain by constructing a series of bivariate logistic regression models. Models to test the risks for each location vs. all other locations also included the 12-Gy volume (V12), Table 2. Post-radiosurgery Location Frontal lobe Cerebellum Temporal lobe Parietal lobe Occipital lobe Basal ganglia Medulla Thalamus Intraventricular Pons Corpus callosum
J. C. FLICKINCER
77.5
et 01.
Table 3. Multivariate
logistic regression analysis of symptomatic postradiosurgery sequelae ___ .-.-__-Variable
-__.
Postradiosurgery injury expression (PIE) score 12-Gy volume (cc) Prior neurological deficit Prior hemorrhage Prior neurological deficit or prior hemorrhage % Isodose Stereotactic MR planning (plus angiography i Number of isocenters Average dose (Gy) in 20 cc .----
--.._--- -.p value
-----
0.0007 0.0081 U.2920 0.6129 0.5940 0.4729 0.5 187 0.7491 0.9758 ..--__-
which was previously shown to correlate with both all PRI changes and symptomatic postradiosurgery seyuelae. As shown in Table 2, locations with similar regression coefficients were grouped together to construct a four-tiered postradiosurgery injury expression (PIE) score for AVM location. PIE scores for AVM locations straddling more than one site were calculated by averaging the PIE scores for each of the involved sites. For exampie, the PIE score for a temporal-occipital AVM is (2 + 3)/2 = 2.5. Our next step was to test the PIE score against other variables of interest in stepwise multivariate logistic regression analysis. Prior multivariate analysis found that maximum dose (D,,,), minimum AVM nidus dose (D&, target dose inhomogeneity, dose rate, and treatment volume had no additional value (p > 0.2) for determining the risk of developing PRI changes when 12-Gy volume was included (5). Although those variables were not retested, we included the number of isocenters, which approached significance in prior analysis (p = 0.07), in the new multivariate analysis along with other previously untested factors. Table 3 lists the results. We could not document any significance is :> 0.5) of the slight trends for reduced symptomatic postradiosurgery sequelae from adding stereotactic MR imaging to conventional angiography, or patients with prior neurological hemorrhage. Prior neurological deficits were not significantly correlated with any decreased risk of symptomatic
injury expression (PIE) classification Regression coefficient* -1.97 -0.88 -0.83 -0.69 +0.43 +0.60 +0.93 + 1.64 +2.03 +2.11 +3.23
2 1.37 i 1.64 r 0.93 -f 0.70 t 0.73 2 1.47 ? 2.23 -+ 0.88 rt 1.90 i 1.35 k 1.75
PIE score 1 2 2 2 3 3 4 4 4 4 4
* Regression coefficient (+ standard error) for bivariate logistic regression models of symptomatic postradiosurgery sequelae for each location that include both 12-Gy volume and the location indicated, as compared to all other locations combined.
% AVM 100
with Symptomatic
Post-Radiosurgery Sequelae .-___-.. -.
Volume (cc) receiving 12 Gy or more Fig. 1. Risk prediction curves for AVM patients derived from multivariate logistic regression analysis that correlate 1%Gy volume to risks of developing symptomatic postradiosurgery sequelae according to PIE (postradiosurgery injury exprewion) score.
276
I. .I. Radiation
Oncology
0 Biology
l
Physics
Volume
40, Number
2, 1998
Table 4. Observed and predicted rates of symptomatic postradiosurgery sequelae by postradiosurgery injury expression (PIE) category PIE: l-l.5 # of patients % Predicted with symptoms % Observed with symptoms
PIE: 1.6-2.5
PIE: 2.63.0
PIE: 3.3-4
50
194
71
17
3.0% 2.0%
7.3% 7.7%
14.6% 12.7%
23.2% 29.4%
postradiosurgery sequelae (p = 0.29, with a trend towards an increased risk). The only two significant independent variables associated with symptomatic postradiosurgery sequelae were PIE location score @ = 0.0007) and 12-Gy volume (p = 0.008). The odds ratios (OR) for symptomatic postradiosurgery sequelae were 3.08 per unit PIE score (95% confidence interval or c.i. = 1.62-5.88) and 1.08 per cc 12-Gy volume (95% c.i. = 1.02-1.14). Figure 1 shows symptomatic postradiosurgery sequelae risk curves (by 12-Gy volume) for different PIE scores from the final logistic regression model. Risks for intermediate fractional PIE scores can be estimated by interpolating between curves for whole integer scores. Table 4 illustrates how closely the predicted risks of symptomatic postradiosurgery sequelae match the observed data for different PIE scores. We hypothesized that areas of the brain more prone to developing symptomatic postradiosurgery sequelae (with high PIE scores) had the same sensitivity to radiation injury (manifested by PRI changes), but differed in whether or not symptoms were expressed. We tested this hypothesis by analyzing the influence of the PIE score on the risk of developing PRI changes in a multivariate logistic regression model that also included the 12-Gy volume. We found that the risk of PRI changes correlated significantly with 12-Gy volume @ = 0.0002, OR: 1.09,95% c.i. 1.04-1.14), but not with PIE score @ = 0.76). Table 5 shows how closely the predicted risks of PRI changes and symptomatic postradiosurgery sequelae match the observed data categorized by 12-Gy volume quartile. Radiation necrosis Because persistent symptomatic postradiosurgery injury or radiation necrosis is the most important clinical endpoint to the patient and physician, we constructed logistic regression models of it (with V12 and PIE score) despite the limited data. The data was analyzed separately in two subsets of the entire database with different patients excluded for actuarial correction (4). PIE score significantly predicted symptomatic necrosis in models from both the first and second data subsets (pl = 0.048, OR, = 2.49, 95% c.i. 1.04-5.95; p2 = 0.005, OR, = 4.69, 95% c.i. 1.88-117).
The risk of symptomatic necrosis increased by an average factor of 3.4 per unit PIE score. Symptomatic necrosis was not correlated with 12-Gy volume @I = 0.18, OR, = 1.055, 95% c.i. 0.98-1.14; p2 = 0.45, OR, = 1.03, 95% c.i. 0.94-l. 13). This corresponded to an average increase in the risk of symptomatic necrosis by a factor of 1.04 per cc of 12-Gy volume. Because 12-Gy volume did not correlate with symptomatic necrosis risk (but did with all postradiosurgery sequelae), we investigated if the resolution rate for symptomatic postradiosurgery sequelae differed according to radiation dose. As shown in Fig. 2, a significantly greater proportion of symptoms resolved in patients with Dmin < 20 vs. 2 20 Gy (89 vs. 36%, p = 0.006). DISCUSSION This study found that symptomatic postradiosurgery sequelae were significantly correlated with both location, as assessed by PIE score, and with 12-Gy volume. We were able to construct logistic risk-prediction curves from this analysis, with dose and volume parameters for the treatment represented by the single parameter, 12-Gy volume. Riskprediction models of symptomatic necrosis, with less data for this endpoint, confirmed similar significant correlation with PIE location score, but 12-Gy volume was consistently less well correlated with symptomatic necrosis. The questions/hypotheses posed at the start of this study can now be addressed as follows: 1. Did the addition of the stereotactic MRI to angiographic imaging for radiosurgery treatment planning reduce the risks of complications? Not appreciably. Our multivariate logistic regression analysis could find no significant correlation (p > 0.5) for reduced risk of symptomatic postradiosurgery sequelae in patients who underwent stereotactic imaging with both MRI and angiography, compared to angiography alone. There are two possible explanations. The first is that we were reasonably successful at excluding normal brain tissue from the radiosurgery targeting using stereotactic angiography alone. The second is that because the risks of complications appear to corre-
Table 5. Observed and predicted rates of postradiosurgery imaging alterations (PRIA) according to 12-Gy volume (V12) quartile
% Predicted with PRIA % Observed with PRIA % Predicted with symptoms % Observed with symptoms
v12: O-3.8 cc
v12: 3.9-7.3 cc
V12: 7.4-11.2 cc
V12: 11.3-42 cc
20.2%
25.4% 24.3% (n = 70) 7.4% 8.5% (n = 82)
32.5% 35.8% (n = 81) 8.4% 14.1% (n = 85)
47.5% 42.4% (n = 73) 14.4%
17.8% (n = 73) 6.1% 3.5% (n = 85)
10.0% (n = 80)
AVM radiosurgery sequelae l % with Persistent Symptomatic Post-Radiosurgery Sequelae 100 90 80 70 60 50 40 30 20 10 !,I,, /,m ,,I,, I,/,, /vs,III1 0IL 6 12 18 24 30 36 42 48 54 0 T I M E (months after symptom onset)
Ijl/I-
Fig. 2. Actuarial persistencerates for symptomatic postradiosurgery sequelae that developed in 30 AVM patients. Patients that received minimum nidus doses (D,,,) < 20 Gy are compared to patients with Dmin z 20 Gy. spond to the total volume of tissue (including the target) irradiated to high doses [instead of just normal tissue at the margin of the target), the contribution to this risk from unintentionally including a small extra volume of normal brain may be minimal. Another similar explanation is that most of the risk for postradiosurgery sequelae is from hemodynamic changes that result from changes that occur to the AVM vessels. This dominant risk would therefore be independent of the volume of surrounding brain treated. Finally, a small decrease in risk with the addition of MR targeting might not have been detected because of the limited size of this study with a limited number of complications. Despite this study’s conclusion that adding stereotactic MR imaging to angiography did not significantly reduce complications, MRI is still essential because of documented problems with inadequate targeting of the AVM nidus with angiography alone (6). 2. Did patients with prior neurological deficits and/orprior hemorrhage have a lower risk of symptomatic postradiosurgery sequelae? Not appreciably. We hypothesized that a hemorrhage cavity surrounding the target tissue would move surrounding tissue farther out from the margin of the treatment volume, reduce the dose of radiation it receives and, possibly, reduce the level of injury response biomodifiers transmitted to this tissue margin from the target tissue. Furthermore, we expected that patients who already had fixed neurological deficits from hemorrhage or prior therapeutic intervention would be less likely to manifest symptoms from radiation injury. Our analysis, however, did not support these hypotheses. We could not document any significant reduction in symptomatic postradiosurgery sequelae for patients with prior hemorrhage (p > 0.6) and prior neurological deficits predisposed toward symptomatic postradiosurgery deficits, if they had any real effect @ = 0.3). In contrast to these findings, Karlsson et al. found lower risks of symptomatic postradiosurgery sequelae in patients with AVM in peripheral locations who had prior
.I. C. FLICKINGER
et al.
1377
hemorrhage, compared to similar patients with no prior hemorrhage (13). Further study is needed of how prior hemorrhage and neurological deficits affect postradiosurgery sequelae that looks at not only incidence, but also the severity and persistence of symptoms. 3. How much does the risk of symptomatic postradiosurgery injury or radiation necrosis vary with different locations within the brain? Risk varies dramatically with location. Prior studies documented that complications were more likely to be seen from treating deep brain AVM because postradiosurgery imaging changes were more likely to be symptomatic if they developed in the brainstem (3, 5. 13). This is the first study to examine the effect of location in further detail. The four-tiered PIE location scale we constructed was highly correlated with the risks of all symptomatic postradiosurgery sequelae and symptomatic necrosis. The risks of both of these endpoints increased by approximately a factor of 3 with each unit step in the PIE scale. An odds ratio of 3 per unit PIE means that, compared with a PIE 1 frontal AVM, similar sized AVM treated to the same dose with PIE scores of 2, 3, and 4 have higher risks of these sequelae by factors of 3, 9, and 27, respectively. Further studies are needed to independently verify the relative rankings of the PIE scale for predicting complication risks in all different possible locations. Quality of life information to help compare complications in different brain locations would also be useful. 4. Does the maximum average dose of 20 cc, as described by Lax and Karlsson (1 I), improve upon or complement logistic risk prediction of postradiosurger? injury with I2-Gy volume? No. Unfortunately, this paper is not an adequate comparison; of the two different methods for risk prediction of AVM radiosurgery. A fair comparison would need to be done in a separate database, rather than data used to derive one of the formulas. In addition, Lax and Karlsson used a different exponential formula instead of a logistic expression. It is fair to state that the maximum average dose of 20 cc (D,,,) did not improve our logistic modeling of symptomatic postradiosurgery injury risk. The p value for D,, in Table 3 for the final multivariate evaluation of this endpoint belies how closely Davg correlates with complications; it only shows how it independently correlates. Davg varies with dose and volume (and therefore complication risk) in a very similar manner as 12-Gy volume. A very slight advantage for a similar term in stepwise multivariate modeling will result in a very high p value for the second factor because, with similar behavior, it does not add to the discriminatory power of the model. Opposite results could potentially be found in a second database with only slight differences. At present, both formulas seem to provide reasonable predictions of the risks of symptomatic postradiosurgery sequelae. and we cannot comment on which is better without further data.
278
I. J. RadiationOncology l Biology 0 Physics
5. Can the risks for permanent symptomatic postradiosurgery injury (radiation necrosis) be predicted by simply multiplying prediction model risks for all symptomatic postradiosurgery sequelae by the overall chance of symptom persistence, or is there some difference in the dose-volume relationship between temporary and permanent injuries? No, they are different.
Data for symptomatic radiation necrosis as an endpoint are harder to come by and, therefore, more difficult to use as an endpoint for analysis. There are fewer events for necrosis to study than with using PRI changes or even symptomatic changes as endpoints, because the majority of injury responses are subclinical (asymptomatic) and the majority of symptomatic cases go on to resolve. Some of these PRI changes may not be radiation injury responses to be avoided but, instead, could represent parenchymal hemodynamic changes or signals from the injured abnormal arteriovenous shunting vessels that result from successful AVM obliteration. We had hoped that findings from studying the larger body of data on all symptomatic postradiosurgery sequelae (temporary and persistent) could be used to directly project the risks for permanent symptomatic injury or necrosis. We found that this would not be accurate in this study. When we attempted to construct logistic risk prediction models for symptomatic necrosis using the factors VI2 and PIE score
Volume 40, Number 2, 1998
(that fitted the data for all symptomatic sequelae), we found a similar correlation with PIE score but no significant correlation with IZGy volume. The latter finding led us to the discovery that a significantly greater proportion of symptoms resolved in patients with Dmin < 20 vs. 2 20 Gy (89 vs. 36%, p = 0.006). Nevertheless, it would be unwise at this time to arbitrarily restrict AVM radiosurgery prescriptions to minimum nidus doses of < 20 Gy because of this finding. Any gain in lower morbidity with lower-dose treatment must be weighed against higher risks of hemorrhage and death from lower obliteration rates (8, 15). Because of the importance of symptomatic necrosis as a tolerance endpoint, reliable risk-prediction models for necrosis are still sorely needed. Because necrosis data in this paper are not even fully adequate for modeling two factors, more data are required for multivariate modeling of all factors, such as Dmin, that could affect this endpoint. In summary, we found that the risks of symptomatic sequelae from AVM radiosurgery can be predicted according to location with the PIE score and by the 12-Gy treatment volume. More study is needed to verify and possibly further refine the PIE location score, as well as to extend these predictions to permanent symptomatic sequelae (symptomatic radiation necrosis) with greater confidence.
REFERENCES 1. Dixon, W. J.; Brown, M. B.; Jennrich, R. I.: BMDP statistical software manual. Berkeley, CA: University of California Press, Berkeley; 1990:1013-1045. 2. Flickinger, J. C. The integrated logistic formula and prediction of complications from radiosurgery. Int. J. Radiat. Oncol. Biol. Phys. 17:879-885, 1989. 3. Flickinger, J. C.; Kondziolka, D.; Kalend, A. M.; Maitz, A. H.; Lunsford, L. D. Radiosurgery-related imaging changes in surrounding brain: multivariate analysis and model evaluation. Radiosurgery 1:229-236, 1996. 4. Flickinger, J. C.; Kondziolka, D.; Lunsford, L. D. Dose and diameter relationships for facial, trigeminal, and acoustic neuropathies following acoustic neuroma radiosurgery. Radiother. Oncol. 41:215-219, 1996. 5. Flickinger, J. C.; Kondziolka, D.; Pollock, B. E.; Maitz, A. H.; Lunsford, L. D. Complications from arteriovenous malformation radiosurgery: multivariate analysis and risk modeling. Int. J. Radiat. Oncol. Biol. Phys. 38:485-490, 1997. 6. Flickinger, J. C.; Lunsford, L. D.; Kondziolka, D.; Maitz, A. M.; Epstein, A.; Simons, S.; Wu, A. Radiosurgery and brain tolerance: An analysis of neurodiagnostic imaging changes following gamma knife radiosurgery for arteriovenous malformations. Int. J. Radiat. Oncol. Biol. Phys. 23: 19-26, 1992. Flickinger, J. C.; Lunsford, L. D.; Kondziolka, D. Radiosurgical dosimetry: Principles and clinical implications. In: DeSalles, A. F.; Goetsch, S. eds. Stereotactic surgery and radiosurgery. Madison, WI: Medical Physics Publishing; 1993:293-306. Flickinger, J. C.; Pollock, B. E.; Kondziolka, D.; Lunsford, L. D. A dose-response analysis of arteriovenous malformation obliteration after radiosurgery. Int. J. Radiat. Oncol. Biol. Phys. 36:873-879, 1996.
9. Kaplan, E. L.; Meier, P. Nonparametric estimation and incomplete observations. J. Am. Stat. Assoc. 53:457-481, 1958. 10. Kjellberg, R.; Hanamura, T.; Davis, K.; Lyons, S.; Butler, W.; Adams, R. Bragg-peak proton-beam therapy for arteriovenous malformations of the brain. N. Engl. J. Med. 309:269-274, 1983. 11. Lax, I.; Karlsson, B.: Prediction of complications in gamma knife radiosurgery of arteriovenous malformations. Acta Oncol. 35:49-55, 1996. 12. Loeffler, J. S.; Alexander, E.; Siddon, R.; Saunders, W.; Coleman, N.; Winston, K. Stereotactic radiosurgery for intracranial arteriovenous malformations using a standard linear accelerator. Int. J. Radiat. Oncol. Biol. Phys. 17:673-677, 1989. 13. Karlsson, B.; Lax, I.; Soderman, M. Factors influencing the risk for complications following gamma knife radiosurgery for cerebral malformations. In: Karlsson, B., ed. Gamma knife radiosurgery of cerebral arteriovenous malformations. Stockholm: Repro Print A. B.; 1996:34-42. 14. Peto, R.; Pike, M. C.; Armitage, P.; Breslow, N. E.; Cox, D. R.; Howard, S. V.; Martel, N.; McPherson, K.; Peto, J.; Smith, P. G. Design and analysis of randomized clinical trials requiring prolonged observation of each patient. Br. J. Cancer 35:1-39, 1977. 15. Pollock, B. E.; Flickinger, J. C.; Lunsford, L. D.; Kondziolka, D. Hemorrhage risk after radiosurgery for arteriovenous malformations. Neurosurgery 38:652-661, 1996. 16. Voges, J.; Treuer, H.; Lehrke, R.; Kocher, M.; Staar, S.; Muller, R. P.; Sturm, V. Risk analysis of linear accelerator radiosurgery. Int. J. Radiat. Oncol. Biol. Phys. 138:10551063, 1996.
ELSEVIER
l
Clinical
Investigation
ANALYSIS OF NEUROLOGICAL SEQUELAE FROM RADIOSURGERY OF ARTERIOVENOUS MALFORMATIONS: HOW LOCATION AFFECTS OUTCOME JOHN
C. FLICKINGER, M.D.,*’
DOUGLAS KONDZIOLKA, M.D.,*” L. DADE LUNSFORD, M.D. *-c
ANN H. MART,
M.S,‘.
‘ANI?
Departments of *Radiation Oncology.+ Neurological Surgery and *Radiology. University of Pittsburgh School of Medicine. Pittsburgh, PA Purpose/Objective: To elucidate how the risks of developing temporary and permanent neuroIogicaI sequelae tram radIosurgery for arteriovenous maIformaHons (AVM) are related to AVM location, the addition of stereotactic magnetic resonance (MR) imaging to angiographic targeting, and prior hemorrhage or neurdogical deficits. Ma&h& and Methods: We evaluated follow-up imaging and clinical data in 332 AVM patients who received gmIIU3 knife radiosurgery at the University of Pittsburgh between 1987 and 1994. All patients had regular clinical or imaging follow-up for a minimum of 2 years (range: 24-96 months, median = 45 months}, There were 83 patients with MR-assisted planning, 187 with prior hemorrhages, and 143 with prior neurological deficits. ResuIts: Symptomatic postradlosurgery sequelae (any neurological problem including headache) developed in 30 (996)f 332 patients. Symptoms resdved in 58% of patients within 27 months with a sigu&a&Iy greater proportion @ = 0.006) resolving in patients with Dmin < 20 Gy vs. 2 20 Gy (89 vs. 36%). The 7-year actuarial rate for developing persistent symptomatic sequelae was 3.8 % . We first evaluated the reIatIve risks for -rent locations to construct a postradiosurgery injury expression (PIE) score for AVM location. Multivariate Iogistic regression analysis of symptomatic postradlosurgery sequeiae identified independent signiIicant correlations with PIE location score (p = O.@tMY7) and 12 Gy volume (p = 0.008), but with none of the other factors tested (p > 0.3). including the add&m of MR targeting, average radiation dose in 20 cc, prior hemorrhage, or ueurok@d deficit. We used these results to construct a risk prediction model for symptomatic postradiosurgery sequelsre. The risk of radiation necrosis was s@i&antIy correlated with PIE score (p < O&48), but not with 12-Gy v&tune. Conclusion: The risks of developing complications from AVM radiosurgery can be predicted according to location with the PIE score, in conjunction with the 12-Gy treatment volume. Further study of factors affecting persistence of these sequelae (progression to radiation necrosis) is needed. 0 1998 Elsevier Science Inc. Radiosurgery, Stereotactic surgery, Arteriovenous
malformation,
Complications, Radiation injury.
studies of injury responses after radiosurgery. In our first AVM study, we were able to significantly correlate postradiosurgery imaging changes (with and without symptoms) only to treatment volume, but not the to integrated logistic formula predictions (6). A second finding was that brainstem location was significantly associated with the development of symptoms in patients developing imaging changes (6). In a second study that also included benign tumors, we documented higher rates of postradiosurgery imaging (PRI) changes with AVM !31%1) than with a meningioma or acoustic tumors (8%), indicating the importance of tissue response from within the radiosurgery target (hemodynamic or otherwise) in contributing to the injury response. Two new tolerance models that described the tolerance of inhomogeneous radiosurgery dose distributions by the volume of a single minimum threshold dose were proposed in this study (3, 7). The logistic threshold-dose vol-
INTRODUCTION Radiosurgery has proven to be a dramatically successful technique for treating arteriovenous malformations by achieving overall high obliteration rates with relatively low morbidity. Dose-selection for radiosurgery mandates balancing the competing risks of higher obliteration rates (and lower risks of death or morbidity from future hemorrhage) for high doses, against lower risks of morbidity from radiation injury with lower doses. The 1% dose-volume isoeffeet line of Kjellberg et al. and the 3% dose-volume isoeffeet curve from the integrated logistic formula for predicting brain radiation necrosis are widely used radiosurgery doseprescription tolerance guidelines that were based on limited data available at the time (2, 10, 12). We now have more clinical data that can be used to refine tolerance predictions. Several important lessons were learned from our prior Reprint requests to: John C. Flickinger, M.D., Joint Radiation Oncology Center, 200 Lothrop Street. Pittsburgh, PA 1.5213.
Accepted for publication 3 I July 1997 273
274
I. J. Radiation Oncology l Biology l Physics
ume model (which used the volume enclosed within a minimum threshold-dose) provided the closest fit to the data (3, 7). Our most recent analysis of imaging changes after AVM radiosurgery revealed several new findings (5). We found that the total volume of tissue receiving 12 Gy or more (the 12-Gy volume or V12) was the only independently significant factor in multivariate logistic regression analysis that could be used to accurately predict the risks of developing all postradiosurgical imaging changes. The resulting model was, therefore, a logistic threshold-dose model with a threshold dose of 12 Gy. Voges et al. used a similar approach to correlate postradiosurgery imaging changes in 56 AVM and 77 tumor patients to a threshold dose of 10 Gy (16). In addition to the correlation of all PRI changes with 12-Gy volume, our study found that symptomatic postradiosurgery imaging changes were correlated with both 12-Gy volume and location, simply classified as either brainstem or nonbrainstem (5). We also found that symptomatic postradiosurgery changes were less likely to resolve than asymptomatic PRI changes (53% vs. 95%). Because of this difference, we concluded that future studies should focus specifically on symptomatic postradiosurgery changes. Several important questions arose, shortly after we completed the last study of AVM radiosurgery complications, that we investigated in this new analysis: 1. Did the addition of stereotactic MRI to angiographic imaging for radiosurgery treatment planning reduce the risk of complications? 2. Did patients with prior neurological deficits and/or prior hemorrhage have a lower risk of symptomatic postradiosurgery sequelae? 3. How much does the risk of symptomatic postradiosurgery injury or radiation necrosis vary with brain location? 4. Does the maximum average dose of 20 cc as described by Lax and Karlsson (11) improve upon or complement logistic risk prediction of postradiosurgery injury with 12-Gy volume? 5. Can the risks for permanent symptomatic postradiosurgery injury (radiation necrosis) be predicted by simply multiplying prediction model risks for all symptomatic postradiosurgery sequelae by the overall chance of symptom persistence, or is there some difference in the dose-volume relationship between temporary and permanent injuries? MATERIALS
AND METHODS
Clinical material and treatment parameters
We evaluated follow-up MRI scans and clinical data in 332 patients who received gamma knife radiosurgery at the University of Pittsburgh between August 1987 and August 1994. Approval for this study was granted by the University of Pittsburgh Institutional Review Board. All patients had
Volume 40, Number 2, 1998
Table 1. Treatment parameters in 332 AVM patients Standard Variable Dmin (GY) Qn,
(GY)
Isodose (%) Isocenters (#) Volume (cc) 12-Gy volume (cc)
Mean
deviation
Range
20.9 37.2 57.4 2.23 4.38 8.47
3.5 7.3 11.1 1.62 3.83 5.97
12-30 22-59 35-90 l-11 0.01-26.3 0.64-41.6
regular clinical or imaging follow-up for a minimum of 2 years (median = 45 months, range: 24-96 months) at the time of data analysis, Thirty-five patients had no MRI exams between 1 and 2 years after radiosurgery, but did have adequate clinical follow-up. They were included for the analysis of symptomatic postradiosurgery sequelae, but were excluded from the analysis of all PRI changes (this left 297 patients with adequate imaging follow-up for analysis). Prior to radiosurgery, AVMs bled in 187 patients. We documented preexisting neurological deficits (not including histories of headache or seizures) the day before radiosurgery in 143 patients from either AVM hemorrhage (n = 116) or prior therapeutic intervention (n = 27). A total of 220 of 332 patients had an AVM-related neurological deficit and/or a history of AVM hemorrhage. Treatment parameters are listed in Table 1. We planned radiosurgery for 83 patients using both stereotactic MR and angiographic imaging. In the other 249 patients, the only stereotactic imaging performed for treatment planning was biplane angiography. We referred to prior nonstereotactic computed tomographic and MR scans for supplemental 3D information to aid shaping of the target volume from the stereotactic biplane angiography. Dose-volume histograms were reviewed to calculate 12-Gy volumes (V12), which we defined as the total volume of all tissue (including the AVM target) receiving an equal or greater radiation dose than 12 Gy. Cumulative dose-volume histograms were also used to calculate the maximum average dose (D,,,) within 20 cc in the tissue around and including the target (11). Statistical analysis
The primary endpoint for statistical analysis was the development of symptomatic postradiosurgery sequelae. This was defined as any new neurological symptoms, including headache, that developed after radiosurgery that could not be attributed to hemorrhage. We also studied the development of all postradiosurgery imaging (PRI) changes as a secondary endpoint. This was defined as the development of new regions of long-TR imaging changes on postradiosurgery follow-up MR imaging, whether or not these were accompanied by symptoms. The final endpoint examined was persistent symptomatic postradiosurgery sequelae or symptomatic brain necrosis. This was defined as the documented persistence of symptomatic postradiosurgery sequelae for more than 2 years after the onset of symptoms. A p value of 0.05 was considered significant for all
AVM radiosurgery sequelae l
statistical tests. Stepwise multivariate logistic regression analysis of the effects of treatment variables on outcome was performed using BMDP software (1). Only variables approaching statistical significance (p < 0.10) were allowed in the final regression models for symptomatic postradiosurgery sequelae and for all PRI changes. We modeled the endpoint of radiation necrosis using a previously described actuarial correction averaging method for logistic regression analysis (4). Actuarial rates for the persistence long-TR imaging changes were calculated using the method of Kaplan and Meier (9). Statistical comparison between actuarial curves was performed with the log-rank test (14). RESULTS Incidence of postradiosurgical sequelae
Symptomatic postradiosurgery sequelae (any neurological problem including headache) developed in 30 (9.0%) of 332 patients. Symptoms resolved in 17 of 30 patients from 4 to 27 months after onset. Symptoms persisted for more than 24 months in 6 of 30 patients (1.8% of all patients), and were still present in 7 of 30 patients with less than 24 months of follow-up after symptom onset. The actuarial rate of symptom resolution was 57.6 ? 11.3% beyond 26 months after onset. The 7-year actuarial rate for developing persistent symptomatic sequelae was 3.8% (9.0% symptomatic X 42.4% unresolved). We identified PRI changes in 90 (30.3%) of 297 patients with regular imaging follow-up. Asymptomatic postradiosurgery imaging changes developed in 21.3% of all patients who had radiosurgery. Multivariate analysis and modeling
We first evaluated the relative risks of symptomatic postradiosurgery sequelae for different locations within the brain by constructing a series of bivariate logistic regression models. Models to test the risks for each location vs. all other locations also included the 12-Gy volume (V12), Table 2. Post-radiosurgery Location Frontal lobe Cerebellum Temporal lobe Parietal lobe Occipital lobe Basal ganglia Medulla Thalamus Intraventricular Pons Corpus callosum
J. C. FLICKINCER
77.5
et 01.
Table 3. Multivariate
logistic regression analysis of symptomatic postradiosurgery sequelae ___ .-.-__-Variable
-__.
Postradiosurgery injury expression (PIE) score 12-Gy volume (cc) Prior neurological deficit Prior hemorrhage Prior neurological deficit or prior hemorrhage % Isodose Stereotactic MR planning (plus angiography i Number of isocenters Average dose (Gy) in 20 cc .----
--.._--- -.p value
-----
0.0007 0.0081 U.2920 0.6129 0.5940 0.4729 0.5 187 0.7491 0.9758 ..--__-
which was previously shown to correlate with both all PRI changes and symptomatic postradiosurgery seyuelae. As shown in Table 2, locations with similar regression coefficients were grouped together to construct a four-tiered postradiosurgery injury expression (PIE) score for AVM location. PIE scores for AVM locations straddling more than one site were calculated by averaging the PIE scores for each of the involved sites. For exampie, the PIE score for a temporal-occipital AVM is (2 + 3)/2 = 2.5. Our next step was to test the PIE score against other variables of interest in stepwise multivariate logistic regression analysis. Prior multivariate analysis found that maximum dose (D,,,), minimum AVM nidus dose (D&, target dose inhomogeneity, dose rate, and treatment volume had no additional value (p > 0.2) for determining the risk of developing PRI changes when 12-Gy volume was included (5). Although those variables were not retested, we included the number of isocenters, which approached significance in prior analysis (p = 0.07), in the new multivariate analysis along with other previously untested factors. Table 3 lists the results. We could not document any significance is :> 0.5) of the slight trends for reduced symptomatic postradiosurgery sequelae from adding stereotactic MR imaging to conventional angiography, or patients with prior neurological hemorrhage. Prior neurological deficits were not significantly correlated with any decreased risk of symptomatic
injury expression (PIE) classification Regression coefficient* -1.97 -0.88 -0.83 -0.69 +0.43 +0.60 +0.93 + 1.64 +2.03 +2.11 +3.23
2 1.37 i 1.64 r 0.93 -f 0.70 t 0.73 2 1.47 ? 2.23 -+ 0.88 rt 1.90 i 1.35 k 1.75
PIE score 1 2 2 2 3 3 4 4 4 4 4
* Regression coefficient (+ standard error) for bivariate logistic regression models of symptomatic postradiosurgery sequelae for each location that include both 12-Gy volume and the location indicated, as compared to all other locations combined.
% AVM 100
with Symptomatic
Post-Radiosurgery Sequelae .-___-.. -.
Volume (cc) receiving 12 Gy or more Fig. 1. Risk prediction curves for AVM patients derived from multivariate logistic regression analysis that correlate 1%Gy volume to risks of developing symptomatic postradiosurgery sequelae according to PIE (postradiosurgery injury exprewion) score.
276
I. .I. Radiation
Oncology
0 Biology
l
Physics
Volume
40, Number
2, 1998
Table 4. Observed and predicted rates of symptomatic postradiosurgery sequelae by postradiosurgery injury expression (PIE) category PIE: l-l.5 # of patients % Predicted with symptoms % Observed with symptoms
PIE: 1.6-2.5
PIE: 2.63.0
PIE: 3.3-4
50
194
71
17
3.0% 2.0%
7.3% 7.7%
14.6% 12.7%
23.2% 29.4%
postradiosurgery sequelae (p = 0.29, with a trend towards an increased risk). The only two significant independent variables associated with symptomatic postradiosurgery sequelae were PIE location score @ = 0.0007) and 12-Gy volume (p = 0.008). The odds ratios (OR) for symptomatic postradiosurgery sequelae were 3.08 per unit PIE score (95% confidence interval or c.i. = 1.62-5.88) and 1.08 per cc 12-Gy volume (95% c.i. = 1.02-1.14). Figure 1 shows symptomatic postradiosurgery sequelae risk curves (by 12-Gy volume) for different PIE scores from the final logistic regression model. Risks for intermediate fractional PIE scores can be estimated by interpolating between curves for whole integer scores. Table 4 illustrates how closely the predicted risks of symptomatic postradiosurgery sequelae match the observed data for different PIE scores. We hypothesized that areas of the brain more prone to developing symptomatic postradiosurgery sequelae (with high PIE scores) had the same sensitivity to radiation injury (manifested by PRI changes), but differed in whether or not symptoms were expressed. We tested this hypothesis by analyzing the influence of the PIE score on the risk of developing PRI changes in a multivariate logistic regression model that also included the 12-Gy volume. We found that the risk of PRI changes correlated significantly with 12-Gy volume @ = 0.0002, OR: 1.09,95% c.i. 1.04-1.14), but not with PIE score @ = 0.76). Table 5 shows how closely the predicted risks of PRI changes and symptomatic postradiosurgery sequelae match the observed data categorized by 12-Gy volume quartile. Radiation necrosis Because persistent symptomatic postradiosurgery injury or radiation necrosis is the most important clinical endpoint to the patient and physician, we constructed logistic regression models of it (with V12 and PIE score) despite the limited data. The data was analyzed separately in two subsets of the entire database with different patients excluded for actuarial correction (4). PIE score significantly predicted symptomatic necrosis in models from both the first and second data subsets (pl = 0.048, OR, = 2.49, 95% c.i. 1.04-5.95; p2 = 0.005, OR, = 4.69, 95% c.i. 1.88-117).
The risk of symptomatic necrosis increased by an average factor of 3.4 per unit PIE score. Symptomatic necrosis was not correlated with 12-Gy volume @I = 0.18, OR, = 1.055, 95% c.i. 0.98-1.14; p2 = 0.45, OR, = 1.03, 95% c.i. 0.94-l. 13). This corresponded to an average increase in the risk of symptomatic necrosis by a factor of 1.04 per cc of 12-Gy volume. Because 12-Gy volume did not correlate with symptomatic necrosis risk (but did with all postradiosurgery sequelae), we investigated if the resolution rate for symptomatic postradiosurgery sequelae differed according to radiation dose. As shown in Fig. 2, a significantly greater proportion of symptoms resolved in patients with Dmin < 20 vs. 2 20 Gy (89 vs. 36%, p = 0.006). DISCUSSION This study found that symptomatic postradiosurgery sequelae were significantly correlated with both location, as assessed by PIE score, and with 12-Gy volume. We were able to construct logistic risk-prediction curves from this analysis, with dose and volume parameters for the treatment represented by the single parameter, 12-Gy volume. Riskprediction models of symptomatic necrosis, with less data for this endpoint, confirmed similar significant correlation with PIE location score, but 12-Gy volume was consistently less well correlated with symptomatic necrosis. The questions/hypotheses posed at the start of this study can now be addressed as follows: 1. Did the addition of the stereotactic MRI to angiographic imaging for radiosurgery treatment planning reduce the risks of complications? Not appreciably. Our multivariate logistic regression analysis could find no significant correlation (p > 0.5) for reduced risk of symptomatic postradiosurgery sequelae in patients who underwent stereotactic imaging with both MRI and angiography, compared to angiography alone. There are two possible explanations. The first is that we were reasonably successful at excluding normal brain tissue from the radiosurgery targeting using stereotactic angiography alone. The second is that because the risks of complications appear to corre-
Table 5. Observed and predicted rates of postradiosurgery imaging alterations (PRIA) according to 12-Gy volume (V12) quartile
% Predicted with PRIA % Observed with PRIA % Predicted with symptoms % Observed with symptoms
v12: O-3.8 cc
v12: 3.9-7.3 cc
V12: 7.4-11.2 cc
V12: 11.3-42 cc
20.2%
25.4% 24.3% (n = 70) 7.4% 8.5% (n = 82)
32.5% 35.8% (n = 81) 8.4% 14.1% (n = 85)
47.5% 42.4% (n = 73) 14.4%
17.8% (n = 73) 6.1% 3.5% (n = 85)
10.0% (n = 80)
AVM radiosurgery sequelae l % with Persistent Symptomatic Post-Radiosurgery Sequelae 100 90 80 70 60 50 40 30 20 10 !,I,, /,m ,,I,, I,/,, /vs,III1 0IL 6 12 18 24 30 36 42 48 54 0 T I M E (months after symptom onset)
Ijl/I-
Fig. 2. Actuarial persistencerates for symptomatic postradiosurgery sequelae that developed in 30 AVM patients. Patients that received minimum nidus doses (D,,,) < 20 Gy are compared to patients with Dmin z 20 Gy. spond to the total volume of tissue (including the target) irradiated to high doses [instead of just normal tissue at the margin of the target), the contribution to this risk from unintentionally including a small extra volume of normal brain may be minimal. Another similar explanation is that most of the risk for postradiosurgery sequelae is from hemodynamic changes that result from changes that occur to the AVM vessels. This dominant risk would therefore be independent of the volume of surrounding brain treated. Finally, a small decrease in risk with the addition of MR targeting might not have been detected because of the limited size of this study with a limited number of complications. Despite this study’s conclusion that adding stereotactic MR imaging to angiography did not significantly reduce complications, MRI is still essential because of documented problems with inadequate targeting of the AVM nidus with angiography alone (6). 2. Did patients with prior neurological deficits and/orprior hemorrhage have a lower risk of symptomatic postradiosurgery sequelae? Not appreciably. We hypothesized that a hemorrhage cavity surrounding the target tissue would move surrounding tissue farther out from the margin of the treatment volume, reduce the dose of radiation it receives and, possibly, reduce the level of injury response biomodifiers transmitted to this tissue margin from the target tissue. Furthermore, we expected that patients who already had fixed neurological deficits from hemorrhage or prior therapeutic intervention would be less likely to manifest symptoms from radiation injury. Our analysis, however, did not support these hypotheses. We could not document any significant reduction in symptomatic postradiosurgery sequelae for patients with prior hemorrhage (p > 0.6) and prior neurological deficits predisposed toward symptomatic postradiosurgery deficits, if they had any real effect @ = 0.3). In contrast to these findings, Karlsson et al. found lower risks of symptomatic postradiosurgery sequelae in patients with AVM in peripheral locations who had prior
.I. C. FLICKINGER
et al.
1377
hemorrhage, compared to similar patients with no prior hemorrhage (13). Further study is needed of how prior hemorrhage and neurological deficits affect postradiosurgery sequelae that looks at not only incidence, but also the severity and persistence of symptoms. 3. How much does the risk of symptomatic postradiosurgery injury or radiation necrosis vary with different locations within the brain? Risk varies dramatically with location. Prior studies documented that complications were more likely to be seen from treating deep brain AVM because postradiosurgery imaging changes were more likely to be symptomatic if they developed in the brainstem (3, 5. 13). This is the first study to examine the effect of location in further detail. The four-tiered PIE location scale we constructed was highly correlated with the risks of all symptomatic postradiosurgery sequelae and symptomatic necrosis. The risks of both of these endpoints increased by approximately a factor of 3 with each unit step in the PIE scale. An odds ratio of 3 per unit PIE means that, compared with a PIE 1 frontal AVM, similar sized AVM treated to the same dose with PIE scores of 2, 3, and 4 have higher risks of these sequelae by factors of 3, 9, and 27, respectively. Further studies are needed to independently verify the relative rankings of the PIE scale for predicting complication risks in all different possible locations. Quality of life information to help compare complications in different brain locations would also be useful. 4. Does the maximum average dose of 20 cc, as described by Lax and Karlsson (1 I), improve upon or complement logistic risk prediction of postradiosurger? injury with I2-Gy volume? No. Unfortunately, this paper is not an adequate comparison; of the two different methods for risk prediction of AVM radiosurgery. A fair comparison would need to be done in a separate database, rather than data used to derive one of the formulas. In addition, Lax and Karlsson used a different exponential formula instead of a logistic expression. It is fair to state that the maximum average dose of 20 cc (D,,,) did not improve our logistic modeling of symptomatic postradiosurgery injury risk. The p value for D,, in Table 3 for the final multivariate evaluation of this endpoint belies how closely Davg correlates with complications; it only shows how it independently correlates. Davg varies with dose and volume (and therefore complication risk) in a very similar manner as 12-Gy volume. A very slight advantage for a similar term in stepwise multivariate modeling will result in a very high p value for the second factor because, with similar behavior, it does not add to the discriminatory power of the model. Opposite results could potentially be found in a second database with only slight differences. At present, both formulas seem to provide reasonable predictions of the risks of symptomatic postradiosurgery sequelae. and we cannot comment on which is better without further data.
278
I. J. RadiationOncology l Biology 0 Physics
5. Can the risks for permanent symptomatic postradiosurgery injury (radiation necrosis) be predicted by simply multiplying prediction model risks for all symptomatic postradiosurgery sequelae by the overall chance of symptom persistence, or is there some difference in the dose-volume relationship between temporary and permanent injuries? No, they are different.
Data for symptomatic radiation necrosis as an endpoint are harder to come by and, therefore, more difficult to use as an endpoint for analysis. There are fewer events for necrosis to study than with using PRI changes or even symptomatic changes as endpoints, because the majority of injury responses are subclinical (asymptomatic) and the majority of symptomatic cases go on to resolve. Some of these PRI changes may not be radiation injury responses to be avoided but, instead, could represent parenchymal hemodynamic changes or signals from the injured abnormal arteriovenous shunting vessels that result from successful AVM obliteration. We had hoped that findings from studying the larger body of data on all symptomatic postradiosurgery sequelae (temporary and persistent) could be used to directly project the risks for permanent symptomatic injury or necrosis. We found that this would not be accurate in this study. When we attempted to construct logistic risk prediction models for symptomatic necrosis using the factors VI2 and PIE score
Volume 40, Number 2, 1998
(that fitted the data for all symptomatic sequelae), we found a similar correlation with PIE score but no significant correlation with IZGy volume. The latter finding led us to the discovery that a significantly greater proportion of symptoms resolved in patients with Dmin < 20 vs. 2 20 Gy (89 vs. 36%, p = 0.006). Nevertheless, it would be unwise at this time to arbitrarily restrict AVM radiosurgery prescriptions to minimum nidus doses of < 20 Gy because of this finding. Any gain in lower morbidity with lower-dose treatment must be weighed against higher risks of hemorrhage and death from lower obliteration rates (8, 15). Because of the importance of symptomatic necrosis as a tolerance endpoint, reliable risk-prediction models for necrosis are still sorely needed. Because necrosis data in this paper are not even fully adequate for modeling two factors, more data are required for multivariate modeling of all factors, such as Dmin, that could affect this endpoint. In summary, we found that the risks of symptomatic sequelae from AVM radiosurgery can be predicted according to location with the PIE score and by the 12-Gy treatment volume. More study is needed to verify and possibly further refine the PIE location score, as well as to extend these predictions to permanent symptomatic sequelae (symptomatic radiation necrosis) with greater confidence.
REFERENCES 1. Dixon, W. J.; Brown, M. B.; Jennrich, R. I.: BMDP statistical software manual. Berkeley, CA: University of California Press, Berkeley; 1990:1013-1045. 2. Flickinger, J. C. The integrated logistic formula and prediction of complications from radiosurgery. Int. J. Radiat. Oncol. Biol. Phys. 17:879-885, 1989. 3. Flickinger, J. C.; Kondziolka, D.; Kalend, A. M.; Maitz, A. H.; Lunsford, L. D. Radiosurgery-related imaging changes in surrounding brain: multivariate analysis and model evaluation. Radiosurgery 1:229-236, 1996. 4. Flickinger, J. C.; Kondziolka, D.; Lunsford, L. D. Dose and diameter relationships for facial, trigeminal, and acoustic neuropathies following acoustic neuroma radiosurgery. Radiother. Oncol. 41:215-219, 1996. 5. Flickinger, J. C.; Kondziolka, D.; Pollock, B. E.; Maitz, A. H.; Lunsford, L. D. Complications from arteriovenous malformation radiosurgery: multivariate analysis and risk modeling. Int. J. Radiat. Oncol. Biol. Phys. 38:485-490, 1997. 6. Flickinger, J. C.; Lunsford, L. D.; Kondziolka, D.; Maitz, A. M.; Epstein, A.; Simons, S.; Wu, A. Radiosurgery and brain tolerance: An analysis of neurodiagnostic imaging changes following gamma knife radiosurgery for arteriovenous malformations. Int. J. Radiat. Oncol. Biol. Phys. 23: 19-26, 1992. Flickinger, J. C.; Lunsford, L. D.; Kondziolka, D. Radiosurgical dosimetry: Principles and clinical implications. In: DeSalles, A. F.; Goetsch, S. eds. Stereotactic surgery and radiosurgery. Madison, WI: Medical Physics Publishing; 1993:293-306. Flickinger, J. C.; Pollock, B. E.; Kondziolka, D.; Lunsford, L. D. A dose-response analysis of arteriovenous malformation obliteration after radiosurgery. Int. J. Radiat. Oncol. Biol. Phys. 36:873-879, 1996.
9. Kaplan, E. L.; Meier, P. Nonparametric estimation and incomplete observations. J. Am. Stat. Assoc. 53:457-481, 1958. 10. Kjellberg, R.; Hanamura, T.; Davis, K.; Lyons, S.; Butler, W.; Adams, R. Bragg-peak proton-beam therapy for arteriovenous malformations of the brain. N. Engl. J. Med. 309:269-274, 1983. 11. Lax, I.; Karlsson, B.: Prediction of complications in gamma knife radiosurgery of arteriovenous malformations. Acta Oncol. 35:49-55, 1996. 12. Loeffler, J. S.; Alexander, E.; Siddon, R.; Saunders, W.; Coleman, N.; Winston, K. Stereotactic radiosurgery for intracranial arteriovenous malformations using a standard linear accelerator. Int. J. Radiat. Oncol. Biol. Phys. 17:673-677, 1989. 13. Karlsson, B.; Lax, I.; Soderman, M. Factors influencing the risk for complications following gamma knife radiosurgery for cerebral malformations. In: Karlsson, B., ed. Gamma knife radiosurgery of cerebral arteriovenous malformations. Stockholm: Repro Print A. B.; 1996:34-42. 14. Peto, R.; Pike, M. C.; Armitage, P.; Breslow, N. E.; Cox, D. R.; Howard, S. V.; Martel, N.; McPherson, K.; Peto, J.; Smith, P. G. Design and analysis of randomized clinical trials requiring prolonged observation of each patient. Br. J. Cancer 35:1-39, 1977. 15. Pollock, B. E.; Flickinger, J. C.; Lunsford, L. D.; Kondziolka, D. Hemorrhage risk after radiosurgery for arteriovenous malformations. Neurosurgery 38:652-661, 1996. 16. Voges, J.; Treuer, H.; Lehrke, R.; Kocher, M.; Staar, S.; Muller, R. P.; Sturm, V. Risk analysis of linear accelerator radiosurgery. Int. J. Radiat. Oncol. Biol. Phys. 138:10551063, 1996.