Table 3. Top 15 risk factor variables for predicting mortality listed in descending order of “importance” by algorithm derived from the training cohort of 376,971 patients.
| Cox model a | Random Forest b | Deep Learning c |
|---|---|---|
| Age | BMI | Smoking |
| Prior diagnosis of cancer | FEV1 | Age |
| Gender | Waist circumference | Prior diagnosis of cancer |
| Smoking | Diastolic blood pressure | Alcohol consumption |
| Prior diagnosis of COPD | Systolic blood pressure | Digoxin prescribed |
| FEV1 | Age | Gender |
| Prior diagnosis of T2DM | Body fat percentage | Warfarin prescribed |
| Prior diagnosis of CHD | Smoking | Townsend deprivation index |
| Diastolic blood pressure | Prior diagnosis cancer | Residential air pollution |
| BMI | Gender | Prior diagnosis of CHD |
| Systolic blood pressure | Skin tone | Statins prescribed |
| Townsend deprivation index | Education | Prior diagnosis of COPD |
| Ethnicity | Prior diagnosis T2DM | Job exposure to hazardous materials |
| MET-min week | Vegetable consumption | Education |
| Education | Fruit consumption | FEV1 |
a ranking determined by strongest to weakest Cox regression coefficients
b ranking determined by largest to smallest mean decreases in accuracy
c ranking determined by largest to smallest scaled importance derived from network weights
orange = top risk factor in all three algorithms; blue = top risk factor in two algorithms; green = top risk factor in one algorithm