IP Requirements:

• Emory IP

Experience Requirements:

• AI/ML Algorithm Development
  

Hyperkalemia (high potassium) is a common in Acute Kidney Injury (AKI i.e. Acute renal failure), Cardiac patients and End Stage Renal Disease (ESRD i.e. Dialysis patients) especially in ICU (6.3%) and associated with cumulative, incremental mortality risks Odds Ratio 3.3 to 7.6, as high as 29.7%.   The biological effects of hyperkalemia manifest on the electrocardiogram (ECG) and may predict mortality.  The mortality effects of EKG changes in hyperkalemia especially in clinically intervenable stages in largely unexplored, current knowledge is limited to extreme ranges of hyperkalemia. The project sponsor/researcher is a nephrologist  and intensivist  with published work and expertise on the topic including prediction algorithms and mortality parameters. 

Aim/Objective: 

Evolve a predictive analytics EKG-based model for mortality in hyperkalemia in the clinically intervenable and modifiable ranges.  Ideally, the predictive model based on explainable AI, supervised learning.

I propose to build on my preexisting work, I already built 2 data bases:

1. a) 12 lead EKG data base which I used for derivation, proof of concept and validation of parameters 
2. b) big data set of wave form data already linked to clinical data which is extracted from MIMIC 4

Please refer to my Journal of Electro cardiology publication for a general idea of hyperkalemia prediction work and EKG mortality parameters. 

Project proposal:

I propose the following sequence utilizing software platforms Python or MATLAB

1. Review of linked wave form MIMIC data for outliers and validation  
2. Filter the MIMIC wave form data with frequency domain analysis, low pass filters
3. R wave and T wave identification (utilize/modify open-source algorithms on MIT/BWH Physionet etc.)
4. Feature extraction of wave form features of clinical/physiological relevance and interest signal averaged T waves (I have published and unpublished data on these parameters to cue the model) –> linear and nonlinear parameters 
5. Extract mortality data from MIMIC along with clinically relevant data (ideally time weighted) that influence mortality 
6. Separate data into derivation and validation subsets 
7. Evolution of predictive analytics mortality model: Supervised learning utilizing wave form features I defer methodology to experts: Broadly propose to use classification methodologies: SVM or Neural networks or decision trees 
8. Run epochs for derivation and validation 
9. Explainability of model probably SHAP 
10. Ideally would use another data set to test the generalizability, validity and robustness of models

Clinical applications and future scope: 

The proposed mortality model would have significant clinical impact and potentially wide scope for further evolution of predictive analytics models. I propose to exploit the model for further evolution of personalized/precision medicine (based on mortality risk) to guide optimal and timely potassium interventions. Such precision medicine is very important in subsets of AKI, Cardiac patients and ESRD. The sponsor has access to multiple academic centers and collaborators in cardiology, nephrology and would like to evolve and validate the model to other real time datasets and expand the scope to monitoring devices, implantable devices and wearables. The project may result in paradigm shifts in clinical care, risk stratification and precision medicine care models for these high-risk patients.