Using machine learning to predict drownings in surf beaches of southwest France

David Carayon

david.carayon@inrae.fr

Bruno Castelle

Eric Tellier

Bruno Simmonet

Jeoffrey Dehez

December 4, 2023

Some context

Should France be your next vacation destination 🏖️ ?

South west France surf beaches

The place to be ! 🌟

Some (😨) numbers

Castelle et al. (2019)

• One of the most dangerous coasts in the world (Castelle et al. 2019) : heavy rip current (Baïnes), shore break

• Surveillance mainly during summer : Thousands of rescues each year

• 20 to 30 fatal drownings each year
  (💪 0 between the flags)

• About \(1\over2\) of swimmers bathe outside of the surveillance zone (Dehez and Lyser 2021)

Can we prevent drownings ? 🤔

Of course, there’s even a world class conference about it !

Can we prevent predict drownings ? 🤔

Spoiler : It was in the title

Previous work in France

• Physical modelling of hazard for South-west France : (Castelle et al. 2019)

• Drowning risk prediction using Log. Regression (Tellier et al. 2021)
• Based on emergency calls database
• Probability of a drowning occuring each day, based on weather, oceanic and crowd data

Goal of this work

• Same philosophy as (Tellier et al. 2021) : Daily prediction based on weather and beach crowd

• Better (and cleaner) data

• New statistical methods 🌟

New statistical methods 🌟

Machine learning 🤖 in a nutshell

AI == ML == statistics & IFELSES

Risk modelling strategy and methods

Let’s play with some data

What’s a risk ?

Our definition :

• Hazard : Is it dangerous ?

• Exposure : Is the beach crowded ?

The predictors

Hazard (daily maximum)

• Wave incidence factor : \({cos}_4H = cos((278 - D\_{SWELL}) \times \frac{\pi}{180})^4\)

• Wave factor : \(HsTp = H_{SWELL} \times P_{SWELL}\)

Exposure

• \(T_{air}\) : air temperature in °C (3-day pred. & daily max value)

• \(day\) : day (ex : \(6\) for \(6^{th}\) of July)

• \(month\) : month (ex : \(7\) for July)

• \(wday\) : weekday (ex : \(1\) for monday)

The outcome

• Water inhalation & respiratory impairements (leads to emergency calls)
• Emergency calls database from 2011 to 2022 | N = 522
• Binary daily data (a drowning occured / no drowning occured)

Technical stuff 🛠️

• Programming language : R
• Reporting and communication : Quarto
• Data wrangling and plotting : {tidyverse} ⭐
• Modelling framework : {tidymodels}
• Reproducible pipelines : {targets} and {renv}

Our challengers

Logistic regression : No tuning parameters

• Classical regression for binary outcome
• Used in previous work (Tellier et al. 2021)

Random Forests : 3 tuning parameters

• # of trees
• # of random variables
• minimum tree depth

XGBoost : 4 tuning parameters

• same as RF
• learning rate
• spoiler: 🏆

Modelling strategy

About these steps

Pre-processing

• Centering, scaling, dummy-coding
• Synthetic Minority over-Sampling Technique (SMOTE) (Chawla et al. 2002) for the outcome
• removing correlations (\(r > 0.9\))

Modelling strategy

About these steps

• Tuning grid generated by Latin Hypercube Sampling (Sacks et al. 1989)

Modelling strategy

Results

360 models later…

Tuning workflow results

I have bad news…

Daily drownings events can’t be predicted with 100% precision.

What would be the best model ?

Improving accuracy : Discretization

• The probability output by the model is discretized using 5 classes :

Testing set (N =600)

Risk class	drownings	No drownings
1	31	502
2	20	44
3	4	20
4	16	11
5	7	8

Discussion

• What is the best drowning prediction model ? Should we prioritize lowering false negatives or false positives ? \(\rightarrow\) Risk management and political decisions

• Low improvements over previous models \(\rightarrow\) dataset limitation ?

• Emergency call database only covers a minority(😱) of all rescues

Perspectives and future work

Perspectives and future work

Want to try this at home ?

Try our 📦 {DrowningPrediction} on and get in touch with us !

Thank you!

Find me online:

david.carayon@inrae.fr

github.com/davidcarayon

@david_carayon

Slides available at dcarayon.fr/slides/WCDP23/

References

Castelle, Bruno, Tim Scott, Rob Brander, Jak McCarroll, Arthur Robinet, Eric Tellier, Elias de Korte, Bruno Simonnet, and Louis-Rachid Salmi. 2019. “Environmental Controls on Surf Zone Injuries on High-Energy Beaches.” Natural Hazards and Earth System Sciences 19 (10): 21832205.

Chawla, N. V., K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer. 2002. “SMOTE: Synthetic Minority over-Sampling Technique.” Journal of Artificial Intelligence Research 16 (June): 321–57. https://doi.org/10.1613/jair.953.

Dehez, Jeoffrey, and Sandrine Lyser. 2021. “Fréquentation des plages océanes et risques de baignade en Aquitaine en 2020. Une étude exploratoire.” Research Report. INRAE. https://hal.science/hal-03549020.

Sacks, Jerome, William J. Welch, Toby J. Mitchell, and Henry P. Wynn. 1989. “Design and Analysis of Computer Experiments.” Statistical Science 4 (4). https://doi.org/10.1214/ss/1177012413.

Tellier, Éric, Bruno Simonnet, Cédric Gil-Jardiné, Marion Lerouge-Bailhache, Bruno Castelle, and Rachid Salmi. 2021. “Predicting Drowning from Sea and Weather Forecasts: Development and Validation of a Model on Surf Beaches of Southwestern France.” Injury Prevention 28 (1): 16–22. https://doi.org/10.1136/injuryprev-2020-044092.