new papers | Notion

Here is a doc with information from the old sources I’ve used:

Title: A ML prediction model to identify risk of firearm injury using electronic health records data

Preventative healthcare and firearm injury, trying to identify the risks of certain patients to firearm injury
Used data from KPSC of patients with injuries from firearms
Identified the type of injury the patient had (fatal, non-fatal, self-inflicted, and non-self-inflicted)
Looked at current or history of suicide attempts and suicidal ideation, history of firearm injury, individual-level socio-demographics, individual-level clinical and healthcare utilization, as well as neighborhood-level information for patients
XGBoost was used to predict all fatal and non-fatal firearm injuries combined as the primary outcome
Used gain metric for feature selection
- They listed the most important features by this metric
They used weird metrics
Results: A probability cutoff that allowed 83% of encounters within 3 years of firearm injuries to be identified (sensitivity) rendered a specificity of 56%. Because of the very low event rate of 0.01%, the number of false positives was high and consequently, the PPV was low (0.0002) at the high-sensitivity probability cutoff
Num features: 170, found 15 important ones
Target: type of firearm injury (including fatal, non-fatal, self-inflicted, and non-self-inflicted injuries)
training (1786 injuries), validation (596 injuries), and test (990 injuries), respectively

Title: Machine Learning Analysis of Handgun Transactions to Predict Firearm Suicide Risk