Insight Health Data Science project - SepRisk: Know Your Sepsis Risk. Save Your Life. (consulting project for Patch'd Medical). This project describes how Cox regression can be used to develop a model to predict the risk of sepsis relative to time,
2. Sepsis: exaggerated
immune system
reaction to infection
1 million affected annually
1/3 don’t survive!
!
! !
!
Annual
healthcare costs
of sepsis:
$24 billion
Sepsis and its impacts
4. Working with the data
Data
• 45K ICU patients
• Sepsis status and time
• Vital signs
• Demographics
• ICD9/10 codes
• Medical history
Pre-Processing
• Vital sign feature
engineering
• Calculated Charlson
Comorbidity Index
• Removed outliers
• Dimension reduction
• Multiple imputation
• Undersampling
Feature Selection
• Univariate Selection
• Non-collinear features
• Significant features
• Interaction terms
5. Data exploration – Relationship between vital signs and sepsis
• Both the LEVEL and VARIABILITY of vital signs were predictive of sepsis
Blood Pressure Blood Pressure
Sepsis occurs when our body has an exaggerated response to infection, thereby leading to severe organ damage and death
1 million people in the US are affected by sepsis every year and almost one-third don’t survive
Sepsis accounts for 24 billion dollars in annual healthcare costs which is a 1.5 billion dollar increase since 2015;
3.5 billion dollars of that, is due to sepsis readmissions.
Sepsis thus takes a massive toll on both the economy and human wellbeing
Worn by patients at home
Seeking a ML model they can integrate into their software – predicts both IF and WHEN sepsis will occur
And identifies relative feature importance.
At present – they have a black box algorithm that only predicts probability of sepsis
Plenty missing data
We can see there is a GREATER PROPORTION OF SEPTIC PATIENTS at higher quintiles of heart rate, and at lower quintiles of blood pressure
Whereas the
And event importance
Which is directly in line with the client’s needs
We have model that can predict sepsis risk with high accuracy, at multiple time points >> comparable to the best published model (using hospital based equipment)
Patch’d believe this will be of tremendous value to their customers using their Patchd device at home
By Pts and Caregivers knowing their sepsis risk relative to time – they can take prompt immediate action to prevent their condition escalating
Worked in the areas of mental health, health policy, applied social research, and ageing
Using data to heal, and make a difference in the world