« Study index
SAVANNAS logo

SAVANNAS study Twitter

Statistical Analysis techniques Versus Artificial Neural Networks for diagnosis and outcome prediction after Acute Stroke

For further information please contact us.
• Chief investigator:
• Nottingham stroke trials office:
Telephone: +44 (0) 115 823 1670
Email:


Title Statistical Analysis techniques Versus Artificial Neural Networks for diagnosis and outcome prediction after Acute Stroke
Acronym SAVANNAS
Chief investigator Professor Philip M Bath
Synopsis Stroke is common, may be mimicked by other conditions, and often has a poor outcome.  We will assess the accuracy of state-of-the-art machine learning with artificial neural network (ANN) models, including deep neural networks (DNN), in the diagnosis of stroke versus mimics, and prediction of early complications (deterioration, stroke recurrence and re-bleeding) and late outcomes (functional outcome, cognition, mood, quality of life, disposition, death) after stroke.  These will be compared with conventional statistical regression models (binary, Cox, ordinal and linear regression).  Neither ANN/DNN nor statistical approaches are used routinely in stroke management in part because of poor accuracy and acceptability to clinicians.  If diagnostic and prediction models are accurate and acceptable to patients and clinicians then their use could be rapidly introduced into clinical care to improve patient management, for example through the use of apps on phones/tablets/computers.  We plan to use individual patient data from acute stroke trials (which will include those with stroke, TIA and mimics) with baseline clinical and imaging, on treatment, diagnostic and outcome information.
Number of patient records Approx. 90,000
Duration 2.5 years
Hypotheses
  1. Mathematical approaches (ANN including DNN, statistical) enhance diagnosis and prediction of stroke and its outcomes
  2. ANN/DNN approaches have superior accuracy to statistical regression approaches for diagnosis and prediction in stroke
  3. Addition of early follow-up information improves outcome prediction
Research questions
  1. Do ANN (including DNN) and/or statistical regression enhance diagnosis of stroke and prediction of outcomes?
  2. Do ANN/DNN have superior accuracy to regression in predicting outcomes (binary, ordinal, continuous) up to days2-10 (death, deterioration, impairment, recurrence, bleeding following thrombolysis, haematoma expansion) and days 90 (death, recurrence, dependency, disability, cognition, mood, quality of life, disposition) after stroke?
  3. Does the addition of early follow-up data, at days 2-10, improve prediction for both ANN and regression?
  4. Does ANN have superior accuracy to regression in diagnosis of stroke?
  5. Does the use of much larger (10-50-fold larger) and more diverse (international/multicentre) datasets improve test statistics, as compared with earlier published smaller datasets?
Analysis methods
ANN/DNN
ANN models will be implemented in R, including the use of cross-validation (to limit over-fitting).  We will test different forms of ANN including feedforward networks, radial basis networks and the latest deep neural network types; and assess whether the additional use of a genetic algorithm optimises ANN models.  Model performance comparison such as comparisons between models, and receiver operating characteristic(area under curve) will also be performed in R.  Checks will be performed using SAS.
Statistical regression models
Regression models used will include binary logistic regression, Cox proportional hazards regression, ordinal logistic regression and multiple linear regression.  All analyses will be run in R, with checks performed using SAS.
Model development and testing
The process will be:
  1. Develop/train the model in a first dataset (using a random selection of approx. 1/3 of the data)
  2. Validate the model in the remaining 2/3 of the first dataset
  3. Test the model in a separate dataset

Current status

Trials shared
with study		 Number of
patients		 Collaborator
ALIAS 1 841 NINDS, M Ginsberg, US
ATTACH-II 1,000 NINDS, AI Qureshi, US
Brain attack study 399 JM Wardlaw, UK
DASH 54 N Sprigg, M Desborough, UK
DCLHb 85 PJ Koudstaal, NL
DEFUSE 3 182 NINDS, G Albers, US
ENOS 4,011 PM Bath, UK
EXTEND-IA TNK parts 1&2 502 B Campbell, Aus
FAST-MAG 1,700 NINDS, J Saver, US
GTN-1/2/3 147 PM Bath, UK
iDEF 294 NINDS, M Selim, US
IST-1 19,435 P Sandercock, UK
IST-3 3,035 P Sandercock, UK
POINT 4,881 NINDS, S. Claiborne, US
RIGHT-1/2 1,190 PM Bath, UK
SO2s 8,003 C Roffe, UK
Stroke mimic datasets 670 W Whitely, UK
TAIST 1,489 PM Bath, UK
TARDIS 3,096 PM Bath, UK
TICH-1/2 2,349 N Sprigg, UK
TriMeth 312 J Dawson, UK
VISTA dataset 32,270 VISTA
Total: 85,945
Contact details
Address: Room S/D2112
Stroke Trials Unit
School of Medicine
University of Nottingham
Queen's Medical Centre
Derby Road
NOTTINGHAM
NG7 2UH
United Kingdom
Telephone:+44 (0) 115 823 1670
Email: