We thank the respondents for providing their case experience and allowing further discussion of this important topic. We would first like to draw attention to specific points in the described case, before discussing some of the more general issues raised.

The respondents report one of many scenarios in which it may be undesirable to use dual antiplatelet therapy in the elective treatment of intracranial aneurysms with flow diverters. In this case, a 47-year-old woman with an 11 mm left ophthalmic aneurysm harbours a significant aspirin allergy. A single Pipeline Shield device under cover of ticagrelor was used to treat the aneurysm. The patient was well at discharge on postoperative day three but then developed symptomatic stent thrombosis on day 6. We draw attention to three points:

1) The respondents state that the stent achieved “perfect wall apposition improved with intra-stent balloon angioplasty.” Setting aside the impossibility of improving “perfect wall apposition” with angioplasty, this does allude to the increasingly understood importance of flow diverter wall apposition. [1] However, digital subtraction angiography assess stent apposition poorly. [1] The use of angioplasty suggests that there may have been some initial concern. Moreover, angioplasty itself may contribute to thrombosis if it promotes activation of the extrinsic clotting pathway by disrupting the endothelial layer. The phosphocholine “Shield” layer reduces thrombosis and platelet aggregation via retarding activation of clotting factor XII at the stent surface(intrinsic pathway). It will exert no effect on the extrinsic pathway should it be activated.

2) The respondents do not mention if platelet response to ticagrelor was tested. This is not a criticism, as it is common to use new generation P2Y12 antagonists without testing their efficacy. However, it is also essential to understand ticagrelor resistance is reported in up to 3% of the population. [2] More recently we have experienced stent occlusions using the aspirin only technique in patients who were subsequently found to be aspirin non-responders. Given the loss of redundancy inherent in a single antiplatelet therapy (SAPT) technique, anti-platelet response testing should be reconsidered.

3) The timing of the symptomatic stent occlusion is noteworthy as it occurred three days post-discharge from hospital. This is somewhat unusual, as typically SAPT stent occlusions occur more acutely. Ticagerol’s shorter half-life means its effect will be subtherapeutic if discontinued for three days. Therefore, given the suspicious timing of the patient's stent occlusion, issues with patient compliance should be considered.

The use of the Pipeline Shield with aspirin only is a technique reserved for extreme circumstances. In aneurysmal SAH patients who cannot be treated successfully without the use of a stent and in whom antiplatelet therapy is undesirable or contraindicated, this technique may be appropriate. This has been stated in the original manuscript and is apparent in the description of the aneurysms treated. Our study reported an asymptomatic thromboembolic complication rate of 14.3% and a symptomatic rate of 7.1%. [3] Given the complex and extremely challenging nature of these cases, a symptomatic thromboembolic complication rate of 7.1% may be acceptable. The respondents point out that the combined rate is 21.4%. Any complications are too many, however, bear in mind that the thromboembolic complication rate in CLARITY was 12.5%. [4] Given the latter patient population underwent routine aneurysm coiling and the former represents patients without conventional treatment options this difference seems acceptable. Moreover, the symptomatic thromboembolic complication rate in the SAPT study is not very different from that reported in either ASPIRe or IntrePED. [5,6]

We believe that the addition of the phosphocholine "Shield" surface modification may allow greater flexibility in the use of antiplatelet therapy. It is increasingly clear that antiplatelet therapy is central to complications with flow diverting stents. Recent meta-analysis demonstrates that a hypo-response to DAPT increases the risk of thromboembolic complications and that a hyper-response increases the risk of haemorrhagic complications. [7] In the setting of acute aneurysm rupture, antiplatelet therapy is even more critical. The use of DAPT in such patients has been shown to increase the risks of haemorrhagic complications dramatically. [8] Therefore, any technology that may reduce our reliance on antiplatelet therapy should be fully explored.

The use of SAPT in complex ruptured aneurysms has nuanced advantages that may not be apparent to all readers. Many operators choose to delay treatment of ruptured aneurysms requiring a flow diverter to avoid dual antiplatelet therapy (DAPT) acutely. The mean time from subarachnoid haemorrhage (SAH) to aneurysm treatment in our study was one day. In comparison to 7 days in the meta-analysis reported by Cagnazzo et al. [9] This is particularly important to bear in mind when comparing retrospective studies as patients who re-ruptured before treatment will be excluded from analysis, however, undoubtedly represent treatment failure. This point is highlighted in a recent report by ten Brick and colleagues. [10] In this study, 44 patients with aneurysmal subarachnoid haemorrhage were treated with flow diverting stents on DAPT in five European centres. A range of flow diverters were used. However, none had the "Shield" surface modification. Aneurysm and patients characteristics were similar to the SAPT series we reported. Aneurysms were treated acutely (mean of three days post-SAH), although not as early as in the SAPT study. [3] Procedure-related complications occurred in 44%. Permanent neurological deficit due to procedure-related complications occurred in 27%. Perhaps of most interest to this discussion, post-procedure intracranial bleeding occurred in 22.7% including five aneurysm re-bleeds, two intraparenchymal haemorrhages not related to ventriculostomy, two ventricular shunt-related haemorrhages and one extra-axial haemorrhage. With extracranial haemorrhagic complications reported in another 9.1%. [10] In the SAPT series, two aneurysm re-bleeds occurred with one patient making a full recovery. The other patient, who was already a poor WFNS grade SAH never regained consciousness from the initial ictus. Both patients were being treated with postoperative heparin infusion at the time of aneurysm re-rupture, a practice which was subsequently abandoned as described in the original manuscript. In spite of similar patient and aneurysm characteristics, the DAPT case series reported good clinical outcomes in 45.5% compared to 64.3% in the SAPT series. Procedure-related mortality occurred in 11.4% and 7.1% of the DAPT and SAPT series, respectively.

Finally, it is essential to clarify the use of the term single antiplatelet therapy (SAPT). We are forced to take considerable responsibility for any misuse of this term. One should not consider the use of only one agent as the goal of this work. Instead, the goal is to offer patients with complex ruptured aneurysms, treatment using flow diverting (or other) stents with as little suppression of platelet function as is possible to avoid thromboembolic complications. The term "single antiplatelet therapy" has begun to appear in the literature, regarding the use of agents such as ticagrelor and prasugrel. Such terminology is literally correct; however, it misses the real goal of this work. As compared to P2Y12 antagonists, aspirin has a much less pronounced effect on platelet function and naturally, less haemorrhagic complication risk. [11] In most patients, agents such as ticagrelor and prasugrel likely suppress platelet function to such a degree that they make the addition of aspirin mostly redundant. [12,13]

We agree that further investigation of this technique is required, as stated in the original article. However, we think a randomised control trial would be inappropriate at this stage. It is somewhat facile to call for a randomised control trial to solve all medical dilemmas. It is safe to say that no treatment option is established for patients harbouring such complex ruptured aneurysms and therefore deciding on a control group would be somewhat arbitrary.

Furthermore, it is likely to be challenging to detect a significant difference in clinical outcomes given the uncertain prognosis of such patients independent of how their aneurysms are secured. Therefore, the number of patients that would need to be randomised likely exceeds 1000. We have elected to take a different approach and will study the Pipeline Shield SAPT technique in a multicentre, single-arm, prospective study. We hope that the first patients will be enrolled before this letter is published. The study will begin in Australia; however, it is planned to expand to other countries, and we encourage experienced and skilled centres such as yours to be a part of this investigation.

Thank you once again for an engaging discussion. Warm regards,

Nathan Manning

References:

​
1 Rouchaud A, Ramana C, Brinjikji W, et al. Wall Apposition Is a Key Factor for Aneurysm Occlusion after Flow Diversion: A Histologic Evaluation in 41 Rabbits. Am J Neuroradiol 2016;37:2087–91. doi:10.3174/ajnr.a4848

2 Warlo EM, Arnesen H, Seljeflot I. A brief review on resistance to P2Y12 receptor antagonism in coronary artery disease. Thrombosis J 2019;17:11. doi:10.1186/s12959-019-0197-5

3 Manning NW, Cheung A, Phillips TJ, et al. Pipeline shield with single antiplatelet therapy in aneurysmal subarachnoid haemorrhage: multicentre experience. Journal of neurointerventional surgery 2019;11:694–8. doi:10.1136/neurintsurg-2018-014363

4 Pierot L, Cognard C, Anxionnat R, et al. Ruptured intracranial aneurysms: factors affecting the rate and outcome of endovascular treatment complications in a series of 782 patients (CLARITY study). Radiology 2010;256:916–23. doi:10.1148/radiol.10092209

5 Kallmes DF, Brinjikji W, Boccardi E, et al. Aneurysm Study of Pipeline in an Observational Registry (ASPIRe). Interventional Neurology 2016;5:89–99. doi:10.1159/000446503

6 Kallmes D, Hanel R, Lopes D, et al. International retrospective study of the pipeline embolization device: a multicenter aneurysm treatment study. American Journal of Neuroradiology 2015;36:108–15. doi:10.3174/ajnr.a4111

7 Ajadi E, Kabir S, Cook A, et al. Predictive value of platelet reactivity unit (PRU) value for thrombotic and hemorrhagic events during flow diversion procedures: a meta-analysis. J Neurointerv Surg 2019;11:1123. doi:10.1136/neurintsurg-2019-014765

8 Hudson JS, Prout BS, Nagahama Y, et al. External Ventricular Drain and Hemorrhage in Aneurysmal Subarachnoid Hemorrhage Patients on Dual Antiplatelet Therapy: A Retrospective Cohort Study. Neurosurgery Published Online First: 2018. doi:10.1093/neuros/nyy127

9 Cagnazzo F, di Carlo D, Cappucci M, et al. Acutely Ruptured Intracranial Aneurysms Treated with Flow-Diverter Stents: A Systematic Review and Meta-Analysis. American Journal of Neuroradiology 2018;39:1669–75. doi:10.3174/ajnr.a5730

10 ten Brinck MF, Jäger M, de Vries J, et al. Flow diversion treatment for acutely ruptured aneurysms. J Neurointerv Surg 2019;:neurintsurg-2019-015077. doi:10.1136/neurintsurg-2019-015077

11 Gresele P. Antiplatelet agents in clinical practice and their haemorrhagic risk. Blood Transfus Trasfusione Del Sangue 2013;11:349–56. doi:10.2450/2013.0248-12

12 KIRKBY N, LEADER P, CHAN M, et al. Antiplatelet effects of aspirin vary with level of P2Y12 receptor blockade supplied by either ticagrelor or prasugrel. J Thromb Haemost 2011;9:2103–5. doi:10.1111/j.1538-7836.2011.04453.x

13 Warner TD, Nylander S, Whatling C. Anti‐platelet therapy: cyclo‐oxygenase inhibition and the use of aspirin with particular regard to dual anti‐platelet therapy. Brit J Clin Pharmaco 2011;72:619–33. doi:10.1111/j.1365-2125.2011.03943.x

We read with interest the response to our manuscript on using machine learning to optimize elderly patient selection for endovascular thrombectomy (1). We acknowledge here, as the author reports, the limitation of SPOT being based on single center data, and the need for multicenter prospective validation of SPOT as next step in development. The author raises additional technical concerns that we do not necessarily view as applicable to this study.

First, we would like to stress the general limitations of artificial intelligence based techniques such as the overfitting and the data specific local optima problems. However, the specific comments brought by the author are not applicable in our case. First, studies on the number of events per predictor are applicable for logistic regressions (LRs) which is not used in the SPOT algorithm. In fact, our results show poor LR performance which is consistent with the rule of thumb of 1 to 10 referred to by the author. Hence, while serving as a good guidance for LR, the rule is not binding and more importantly it does not guarantee the generalization of the learned model. To further illustrate, classification models using convolutional neural networks have millions of parameters and are trained with datasets that, in most cases, do not have millions of samples in each group. However, these models have acceptable generalization capabilities and are tested using the data-split method. In SPOT, the model at its core is a regression model with continuous output. More importantly, while the overfitting concern is a valid one with the high area under a curve, SPOT was tested in an adequate fashion using data-splitting, a well-accepted validation scheme to test the generalization capability of the model, and thus detect overfitting. In fact, studies discussing events per variable rule for logistic regression use data-splitting as one of the validation method (2). The approach used for testing SPOT using a prospective data not part of the training is the most stringent approach to test a model. In fact, state of the art machine learning model evaluations use the data-splitting technique, and consider the performance of the testing data as the golden metric to judge upon the model’s generalization and over-fitting (3,4).

While we agree with the fact that machine learning are data hungry, the size of the dataset is highly dependent on the problem at hand. For example, in sparse regression models, the number of samples in the training dataset can be orders of magnitude smaller than that of the parameters. However, sparse regression models joined with proper training techniques are able to generalize to unseen data (5). Again, this ability is tested using the data-split method which was used to test SPOT.

Further, we emphasize that although SPOT returns a continuous output of mRS scores, the tool will specifically report grouped outcomes into mRS 0-2 and mRS 3-6. The decision to choose this dichotomy of outcomes was to be consistent with clinical trials that predominantly report functional independence as outcome measure to guide interventions even when elderly patients were included. In response to the concern about returning probability for outcomes, and as stated in the manuscript, when SPOT returns poor outcome prediction, its negative predictive value was 95.2% which represents the probability for a patient that screened negative to have a poor outcome and this probability is reported in the text. In the current form, SPOT does not return a probability for every mRS score.

Finally, while we agree that multicenter data is needed to additionally validate SPOT as a tool as stated in the manuscript, the current version of SPOT does meet technical requirements for a validated tool. We do stress again that “SPOT is designed to aid clinical decision of whether to undergo ET in elderly patients” (1), and not a stand-alone tool.

References
1. Alawieh A, Zaraket F, Alawieh MB, Chatterjee AR, Spiotta A. Using machine learning to optimize selection of elderly patients for endovascular thrombectomy. J Neurointerv Surg. 2019.
2. Harrell FE, Jr., Lee KL, Mark DB. Multivariable prognostic models: issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors. Stat Med. 1996;15(4):361-87.
3. Kohavi, Ron. "A study of cross-validation and bootstrap for accuracy estimation and model selection." Ijcai. Vol. 14. No. 2. 1995.
4. Arlot, Sylvain, and Alain Celisse. "A survey of cross-validation procedures for model selection." Statistics surveys4 (2010): 40-79.
5. Donoho, David Leigh, et al. Sparse solution of underdetermined linear equations by stagewise orthogonal matching pursuit. Department of Statistics, Stanford University, 2006.