Case Studies
Brainome allows multiple industries to save time and money with optimized machine learning models that are human readable. Our measure-first approach allows you to estimate costs prior to building models. Read our case studies or get in touch to create a case study with us.
Use Case: Stock Market Prediction
Problem:
A hedge fund used Brainome to enhance their prediction model.
The Data:
2000+ equities were placed into 3 buckets (big winners, big losers, everything else).
Predictive models were generated from the fund’s proprietary data set spanning 7 years of market data.
Result:
Using our predictions in combination with their own, the fund was able to improve accuracy by up to 25%, with results generated in only hours.
Contact us to learn more about Brainome for Stock Market Prediction
Use Case: Theoretical Physics
Problem:
Rework of 2014 model published in Nature Communications: “Searching for Exotic Particles in High-energy Physics with Deep Learning”
The Data:
5 million instances; 18 features; 2 classes
Result:
Contact us to learn more about Brainome for Theoretical Physics.
Use Case: IoT
Problem:
Video infrastructure client needs very small models for mobile devices
The Data:
16K instances; 107 features; 2 classes
Result:
Both Brainome and sklearn produce models that achieve 85% accuracy but Brainome’s model is 100 Kilobytes in size whereas the sklearn model is 20 Megabytes (200X larger).
Contact us to learn more about Brainome for IoT.
Human Microbiome Project: Identify Disease Markers
Problem:
Identify predictors of inflammatory bowel disease
The Data:
1,627 instances; 71,266 features; 2 classes (healthy vs IBD)
Result:
In 15 minutes* BTC produces a random forest that achieves 99% true-positive-rate for IBD & 94% true-positive-rate for Healthy in a 100 Kilobyte model.*
200 predictive features identified
Human Microbiome Project: https://portal.hmpdacc.org/
*Data set analyzed using Dell XPS-13 laptop, Intel i7 CPU 3 GHz, 32 GB RAM, 2 TB SSD
GTEx Tissue Classification
Problem:
Classify tissue types based on RNA seq-data
The Data:
16,417 instances; 56,200 features (genes); 21 classes
Result:
In 1.5 hours, BTC produces a model that achieves 98.69% accuracy and is only 478 Kilobytes in size.*
273 predictive features identified
GTEx data: https://www.gtexportal.org/home/datasets
*Run on 48 Core AMD Ryzen 3960X, 128 GB RAM
Cancer Genome Atlas
Problem:
Identify genes that are over-expressed in different tumor types
The Data:
21,000 gene expressions for patients with 33 different types of cancer and a sample set of healthy patients combined into a 11,000 row, 21,000 column, 34 class table
Result:
Brainome model achieves 90% accuracy in predicting cancer type based on over expression of specific genes
Cancer Atlas: https://portal.gdc.cancer.gov/
GTEx (healthy): https://www.gtexportal.org/home/datasets
