Investigators
Dr. Oliver Armitage, Dr. Pascal Fortier-Poisson, Dr. Lorenz Wernisch, Antonin Berthon, Tristan Edwards
Contact info (email)
oearmitage@gmail.com
1. Define context(s)
identify/explore new therapies
reveal new biological insights
Current Conformance Level / Target Conformance Level
Adequate
Primary goal of the model/tool/database
For the wealth of pre-clinical studies investigating bioelectronic control of the heart, there is limited information and characterisation of the neural circuits targeted by neuromodulation therapies. We will build a two-part machine learning method to analyse the vagus nerve dynamics as key parameters of stimulation therapy are varied: stimulation current, pulse width, frequency, and duty cycle. The first part of the method will model cardiovascular and neurological responses elicited by stimulation parameters and identify therapeutically valuable parameter sets. The second part of the method will isolate segments of spontaneous of vagus nerve activity that encode changes end-points used to evaluate the efficacy of neuromodulation treatments for heart failure. The method aims to demonstrate how neuronal recordings can be used to adjust neurostimulation parameters and how this modelling paradigm can be applied to optimise neuromodulation treatments in other therapeutics areas, given suitable datasets.
Biological domain of the model
Chronic Heart Failure
Structure(s) of interest in the model
Right and Left Vagus Nerve, Heart
Spatial scales included in the model
10^-1 – 10^-4m
Time scales included in the model
10^2 – 10^-5 s
Other uses for the model (optional)
Model can also be used for better calibrating existing or currently developed therapies
2. Data for building and validating the model
| Data for building the model | Published? | Private? | How is credibility checked? | Current Conformance Level / Target Conformance Level |
|---|---|---|---|---|
| in vitro (primary cells cell, lines, etc.) | ||||
| ex vivo (excised tissues) | ||||
| in vivo pre-clinical (lower-level organism or small animal) | Yes (Ward) | No | Peer reviewed paper published and available on SPARC | Extensive |
| in vivo pre-clinical (large animal) | Yes (BIOS) | Data is stored privately, but has been used to publish methods in medical (INS) and computational conferences (NeurIPS) | Neural data is verified through extensive feature analysis designed to distinguish neural signals from other bioelectronical signals from the body (ECG, EMG) or other sources of noise and artefacts (external, hardware). | Adequate |
| Human subjects/clinical | ||||
| Other: ________________________ |
| Data for validating the model | Published? | Private? | How is credibility checked? | Current Conformance Level / Target Conformance Level |
|---|---|---|---|---|
| in vitro (primary cells cell, lines, etc.) | ||||
| ex vivo (excised tissues) | ||||
| in vivo pre-clinical (lower-level organism or small animal) | ||||
| in vivo pre-clinical (large animal) | Yes (Vaseghi & BIOS) | Vaseghi: No; BIOS: Data is stored privately, but has been used to publish methods in medical (INS) and computational conferences (NeurIPS) | Vaseghi: Published peer reviewed paper and available on SPARC. BIOS: Stimulation parameter ranges are based on literature (Ardell et al.), we repeat measurements of the response to each stimulation setpoint and assess consistency of the recordings through multiple recording channels. | Extensive |
| Human subjects/clinical | ||||
| Other: ________________________ |
3. Validate within context(s)
| Who does it? | When does it happen? | How is it done? | Current Conformance Level / Target Conformance Level | |
|---|---|---|---|---|
| Verification | BIOS Software QA | Throughout development of the acquisition software an independent review process is used prior to release | There is an assessment of software risk and line by line review, regression testing and global testing to insure code meets conceptual implementation | Extensive |
| Validation | Machine Learning Researcher/Developers | Throughout development and during in-vivo trials | Models are validated both on online and offline. Online validation is done during BIOS internal in-vivo trial on porcine subjects. Offline validation is done on public (Ward) and internal (BIOS in-vivo trials) datasets. Such datasets are used to create simulators mapping stimulation parameters to neurological and physiological responses, which are then used to evaluate and validate the models in a simulated online setting. Also, standard model training practise includes splitting available data to training and test sets to ensure that model performance can be tested on data unseen by the model (we normally use k-Fold cross validation). | Extensive |
| Uncertainty quantification | Machine Learning Researcher/Developers | When selecting the appropriate model to use and when training and testing the model | The method used to relate stimulation parameters to physiological response inherently estimates uncertainty as it is a probabilistic model. We calculate the classification accuracy in the prediction of cardiac activity. | Extensive |
| Sensitivity analysis | Machine Learning Researcher/Developers | After model training and testing on the validation set, but before the model is then used in practice | We test the model on multiple subjects to assess robustness to inter-patient variability, across a diverse range of input spaces and objective functions. Each model is tested with inputs that cover the range of expected inputs and also with a set of inputs outside the range of expected inputs. The model sensitivity to variation in the inputs is tested by running it on a range of inputs. | Extensive |
| Other:__________ | ||||
| Additional Comments |
4. Limitations
| Disclaimer statement (explain key limitations) | Who needs to know about this disclaimer? | How is this disclaimer shared with that audience? | Current Conformance Level / Target Conformance Level |
|---|---|---|---|
| Assume that the relationship between the vagal nerve activity and response to stimulation parameters can be effectively modelled with continuous functions | Computational Neuroscientists, Neuromodulation Device Manufacturers, Cardiologists | Modelling and processing shared in Jupyter notebook that will be comprehensively commented, including analysis of this phenomena. Parameters of the model can be varied by users of OSPARC to test limits of modelling approach. | Comprehensive |
| Assumed when training the model to identify aspects of vagal nerve recordings which encode cardiac changes that all variability originates from neural communication where other sources of variability or modulation exist (neural plasticity, autonomic remodelling, inflammatory response) | Experimental Neuroscientists, Computational Neuroscientists, Hardware developers | Assumptions will be stated in model publication, and alongside visualisations on OSPARC | Comprehensive |
5. Version control
| Current Conformance Level / Target Conformance Level |
|---|
| Comprehensive |
| Naming Conventions? | Repository? | Code Review? | |
|---|---|---|---|
| individual modeler | Yes | Yes | Yes |
| within the lab | Yes | Yes | Yes |
| collaborators | Yes | Yes | Yes |
6. Documentation
| Current Conformance Level / Target Conformance Level | |
|---|---|
| Code commented? | Extensive: BIOS follow ISO13485 and have internal SOP for code quality and release workflow |
| Scope and intended use described? | Extensive: Doc Strings + industry standard saved model (TensorFlow) formats |
| User’s guide? | Extensive: Jupyter Notebook + Read Me |
| Developer’s guide? | Extensive: Wiki + Human readable config |
7. Dissemination
| Current Conformance Level / Target Conformance Level |
|---|
| Extensive |
| Target Audience(s): | “Inner circle” | Scientific community | Public |
|---|---|---|---|
| Simulations | Yes | SPARC | No |
| Models | Yes | SPARC | Yes |
| Software | Yes | No | No |
| Results | Yes | SPARC | Yes |
| Implications of results | Yes | SPARC | No |
8. Independent reviews
| Current Conformance Level / Target Conformance Level |
|---|
| Extensive |
| Reviewer(s) name & affiliation: | Guillaume Lajoie, Assistance Professor at Université de Montrèal |
|---|---|
| When was review performed? | Monthly evaluations of modelling strategy |
| How was review performed and outcomes of the review? | Modelling approach and results were presented and evaluated monthly during development of simulation. Code review to be performed prior to submission of computational node on OSPARC platform. |
9. Test competing implementations
| Current Conformance Level / Target Conformance Level |
|---|
| Adequate |
| Yes or No (briefly summarize) | |
|---|---|
| Were competing implementations tested? | Yes - Stimulation Mapping: evaluated on multiple therapeutic areas Yes - Vagus Nerve decoding: when operating in low data regime, deep networks will struggle to learn key input features. For this reason and ease of dissemination, basic models will be tested first, and predictive performance will be compared. |
| Did this lead to model refinement or improvement? | Yes – Stimulation Mapping: compared use of GPs to basic statistical methods. GPs gave us the flexibility to combine other preprocessing steps (detrending ECAPs, combining study cohort data) No – Vagus Nerve Decoding: at current stage basic model has been developed to benchmark deep learning models |
10. Conform to standards
| Current Conformance Level / Target Conformance Level |
|---|
| Extensive |
| Yes or No (briefly summarize) | |
|---|---|
| Are there operating procedures, guidelines, or standards for this type of multiscale modeling? | BIOS engineering developments are performed within an ISO13485 compliant framework. Internally we have a number of SOPs, selected relevant procedures include: - Design Control - Document Control - Agile Sprint Process - Coding Standard Relevant standards include: - AAMI TIR45-12: - ISO 14971:2007 - IEC 62366 |
| How do your modeling efforts conform? | All software development efforts are required to conform with our SOPs. Our Machine Learning team, falls under this jurisdiction. |