(23) BEYOND ONE-SIZE-FITS-ALL: A MACHINE LEARNING APPROACH TO CLASSIFYING FIRE DEPARTMENT PHYSICAL FITNESS NEEDS

Abstract Details:

Introduction: Firefighters execute physically demanding tasks in extreme conditions. However, many firefighters exhibit suboptimal levels of physical fitness (PF), leading to diminished performance in occupational tasks and elevated health risk. Evidence supports that PF levels among firefighters is heterogenous and personalized training is more effective than generic training, indicating a shift away from ‘one-size-fits-all’ is necessary. Many fire departments incorporate annual PF measures into health and wellness programs. The efficacy of such programs can be enhanced by recognizing variations in individuals’ PF levels across a department. To support practitioners in efficiently identifying subgroups in fire departments, machine learning (ML) algorithms are a potentially useful option.

Purpose: The purposes of the study were to 1) employ an unsupervised ML technique to identify distinct subgroups among firefighters and 2) examine differences in PF levels and demographic characteristics between the identified subgroups.

Methods: 1406 firefighters’physical fitness records from 2022 wihin a single fire department were retrospectively analyzed. Physical fitness measures included body fat percentage (BF%), muscular fitness (maximum pull-up, sit-up, and push-up repetitions) and estimated aerobic capacity (VO2max) via a 3-minute step test. K-means cluster (KMC) analysis was used to identify subgroups of firefighters based on PF measures. The number of KMCs was determined visually using the elbow method. Analysis of variances and chi-square tests were conducted to assess differences between the groups identified from the KMC analysis.

Results: The KMC analysis identified 4 groups of firefighters (Table 1). As expected, there were significant differences (p< 0.001) in PF measures between the 4 groups. One group (Cluster 4) scored higher than all other groups on PF tests, except for estimated VO2max (Cluster 1 >Cluster 4). Another group (Cluster 2) had the lowest scores on all muscular PF tests and the highest BF%. Significant differences (p< 0.001) in age, years of service, sex and race distribution were observed between groups.

Conclusion: Overall, KMC analyses revealed distinct subgroups among firefighters, providing insight into the heterogeneous nature of firefighter PF within one department. Notably, while group-level differences in age and sex were evident, individual-level data revealed that females and older firefighters may belong to more 'fit' subgroups. Thus, ML algorithms provide a means to unbiasedly assess the existence of subgroups and determine the composition of each group, thereby offering a data-driven approach for optimizing the targeting of PF interventions to specific subgroups and individuals.

PRACTICAL APPLICATIONS: By leveraging ML, fire departments can employ new approaches to enhance the delivery of their PF initiatives, ultimately supporting the  occupational preparedness, health, and wellness of their firefighters.

Acknowledgements: None