Screening Smarter
Comparing machine learning models of varying size and complexity for title-and-abstract screening of public health literature
- Senior Honors Thesis
- Spring 2027
Abstract
Systematic reviews are the backbone of evidence-based public health, but screening thousands of titles and abstracts by hand is slow and costly. Machine learning can help, yet the trend toward ever-larger models brings rising compute costs and energy use. This thesis compares machine learning techniques of different sizes and complexities, from a lightweight bag-of-words regression classifier and JEV encoding to four models spanning a range of parameter sizes, to determine which offers the best balance of screening performance, cost, and efficiency.
Research questions
- 1How do simple classifiers compare to larger models at identifying relevant public health papers?
- 2Does increasing model size meaningfully improve screening performance?
- 3Which approach gives the best trade-off between accuracy, cost, and energy use?
Approach
Baselines
A bag-of-words regression classifier and JEV encoding serve as simple, transparent, low-cost approaches.
Scaling up
Four models of increasing parameter size test whether added complexity pays off for screening.
Evaluation
Models are compared on screening performance alongside practical costs like compute time and energy.
At a glance
6
Approaches compared, from bag-of-words to large models
4
Model parameter sizes tested head-to-head
1
Goal: the most effective model that is also the most sustainable