Fire Alarm Prediction

Improving the Fire Alarm Program and Emergency Service Efficiency

Ming Cao, Isabelle Li, Luciano Lu, Sihan Yu

Introduction

• Problem Statement
  
• Why this matter
  
• Goal

Data Sources

Source	Description	Key Variables / Features
Fire Incidents (NFIRS-coded)	Incident points (2024–2025)	Label: 1 = 7xx false alarm; 0 = 1xx/2xx; time + geometry
L&I Building Certifications (OpenDataPhilly)	Building inspection/cert records (incl. fire alarm systems)	fire_alarm_status (Active/Expired/Deficient); tract-level active/expired rates
Fire Stations (OpenDataPhilly)	Station locations	distance / near_station_200m
ACS (tidycensus)	Tract socioeconomic context	median_income, ba_rate, unemployment_rate, black_share, total_pop
Weather (METAR via riem, PHL)	Hourly conditions (2024–2025)	temp/wind/precip + timestamp join

False Alarm: Scale + Space

Time + Possible mechanism

Methods Overview

• Technique: Logistic regression (logit)

• Why appropriate: interpretable, outputs probabilities, easy to operationalize

• Validation: k-fold cross-validation, ROC/AUC

• Threshold choice: balances false positives vs false negatives

Results & Performance

Model	Feature set (summary)	AUC
Logit Model 1	Baseline (operational / weather)	0.554
Logit Model 2	+ Neighborhood context (ACS tract)	0.670
Logit Model 3	Full model (+ built env / certificates / proximity)	0.672

Operating Threshold & Risk Tiers

Model 3 (predicting 1 = false alarm):

• Conservative threshold = 0.9:
  Sensitivity = 0.08, Specificity = 0.96 (FP = 493, FN = 42,024)
  
• Aggressive threshold = 0.4:
  Sensitivity = 1.00, Specificity = 0.00 (FP = 13,571, FN = 23)

Important: The model is for prevention and planning — not to delay or deny emergency response.

Recommendations

• Operational: Tiered response package
  
  • Smaller crew + fewer apparatus for likely false alarms
    
  • Bring education materials + basic inspection tools

• Long-term:
  
  • Strengthen compliance
    
  • Community education

Limitations & Next Steps

• Limitations
  
  • Data
    
  • Model
    
  • Bias & harm risk
• Future Possibilities

Questions?

Thank You!

Contact: isabli@upenn.edu · GitHub: [https://github.com/Isabelliiii]

Contact: yusihan@upenn.edu · GitHub: [https://github.com/sihan-yu429]

Contact: lu25@upenn.edu · GitHub: [https://github.com/lluluciano0505]

Contact: cao4@upenn.edu · GitHub: [https://github.com/bananafish9107]