How Data Science Powers Insurance Telematics

Introduction

In this episode of the Dolphin Technologies podcast, Harald speaks with Katharina, Chief Data Scientist at Dolphin Technologies,
about how data science, machine learning, and AI power the core functions of insurance telematics. From detecting risky driving
to anticipating accidents before they happen, the discussion offers a deep dive into how raw data is transformed into actionable
insights for better decision-making.

The Role of Data Science in Telematics

Data science plays a pivotal role in converting smartphone sensor data into meaningful outputs that support both insurers and end users. At its core, data science enables:

• Accurate trip classification (e.g., identifying transport mode)
• Real-time processing on the device for privacy and efficiency
• Actionable interpretation of movement patterns
• Integration of AI to automate insights and decision-making

Combining Methodologies: Heuristics and Machine Learning

Rather than relying on one method, Dolphin uses a hybrid approach:

• Heuristics: Rule-based logic like “buses stop more frequently” to quickly classify transport modes.
• Machine Learning: Identifies complex patterns in noisy data, such as detecting if someone is using a phone while driving.
• Supervised and Unsupervised Learning: Supervised models learn from labeled examples; unsupervised models identify user segments or behaviors without labels.

Detecting Risky Driving Behavior

One of the most important use cases is understanding how—and how safely—someone drives. This includes:

• Detecting smartphone usage during trips (calls, texting, scrolling)
• Classifying phone interaction using gyroscope and accelerometer data
• Training models on large-scale labeled datasets
• Ensuring efficient data use by processing data locally on the device

Crash Detection with AID (Automatic Impact Detection)

Dolphin’s crash detection system listens for specific acceleration patterns to identify vehicle impacts. Key features include:

• Real-time analysis of accelerometer data
• Recognition of crash-specific movement signatures
• Model training with both crash and non-crash data to reduce false positives

Predictive Modeling for Accident Prevention

Looking beyond detection, predictive modeling helps anticipate future risk. Data science supports:

• Forecasting accident-prone scenarios based on behavior and location
• Sending personalized warnings (e.g., weather alerts)
• Using behavioral data to proactively prevent incidents

AI and Generative Capabilities

AI enhances communication and automation:

• Generative AI writes context-aware messages (e.g., weather warnings)
• Predictive analytics determine the optimal time for engagement
• Multilingual and tone-adaptive content delivery

Summary

Data science and AI aren’t just back-end tools—they are essential to delivering safer, smarter, and more personalized telematics services. By combining heuristics, machine learning, and predictive models, Dolphin Technologies provides:

• Improved trip analysis and behavior scoring
• Real-time crash detection
• Risk prevention through smart forecasting

Insurance telematics is evolving from reactive scoring to proactive risk management, thanks to data-driven technologies.

Want more insights? Subscribe to the Dolphin Technologies podcast for expert discussions on smart mobility, telematics, and digital innovation in insurance.

---

Harald
Hello and welcome. This is Harald this time on this side of the table and not Yasmin, because we have a very special guest today. Welcome Katharina.

Katharina
Thanks for having me.

Harald
We just discussed that this is actually your first appearance in a podcast and so it’s your first experience as a podcast guest. And as I can say so far, you’re doing an awesome job.

Katharina
Woohoo! Yeah. Great to hear.

Harald
So, Katharina is our chief data scientist, and today we want to talk about how we use data science in insurance telematics. And we have Katharina to tell us how we do that.

Katharina, we have our weekly calls and you always excite me with your knowledge in data science. We also teach together in university, data science, which is strange because I’m not a data scientist, but I do the business part and you do the technical part. So, I think the most important thing to learn these days is: what’s the difference between data science, machine learning models, and AI?

Katharina
Well, that’s a good one. Let’s briefly start with giving a short overview of data science.

Data science basically is how we turn data into understanding and making smarter decisions. I think of it as a toolbox with specialized instruments to help you through the data process. It supports gathering relevant data, making meaningful analysis, visualizations, reporting, and interpretation so you can take smart data-driven decisions.

AI is also a set of tools and algorithms, but they support machines in doing the decision-making for us. The goal is that the machine can mimic human-like decision making.

Harald
That’s awesome. In a previous episode, we talked about artificial intelligence and how we use it in insurance telematics. Today I would like to discuss how we use data science, machine learning, and other methodologies in insurance telematics.

Katharina
Let me answer this by giving you specific examples of our services because our whole SDK environment is built upon data science.

Harald
SDK is our software development kit. This is the telematics foundation you’re talking about.

Katharina
Yes, exactly. And we employ data science and AI tools throughout the whole process. It starts on the device itself, how we process and collect the data, and ends with how we interpret it.

One very important interpretation layer is trip classification, where we try to understand which mode of transport you’re using. When we detect movement, it’s not immediately clear how you’re moving.

Harald
So when you say how I’m moving, you’re referring to whether I’m using a car, public transport, a train, tram, bus, or whatever.

Katharina
Exactly. Or whether you’re walking or riding a bike. There are different layers of complexity to answer this. We do this mostly through smartphone-based sensors: GPS, speed, accelerometer, gyroscope.

Especially distinguishing between being in a car or on a bus is most complex. For instance, buses usually stop at set intervals while a car would not. We use this type of domain knowledge to interpret the data.

Harald
So what you’re saying is we combine different data science methodologies: machine learning to learn from patterns, and heuristics or rule-based methods like “buses stop more frequently.”

Katharina
Exactly. It’s about complexity and system efficiency. We don’t throw all data into algorithms. We look for strong signals. If a signal is strong, we can build a heuristic. For nuanced patterns, we use machine learning.

Harald
Now we talked about detecting mode of transport, which is essential because if we want to score driving behavior, we need to know if it’s a reckless bus driver or yourself. Where else do we implement machine learning and data science in our processes?

Katharina
Risky driving and driver behavior scoring is key. A major aspect is interaction with the smartphone while driving.

Harald
So distracted driving.

Katharina
Exactly. Phone calls are easy to detect—OS tells us. But texting, reading, scrolling—especially on iOS—is harder.

We built a machine learning classifier to understand if the device is being used during a trip using accelerometer and gyroscope data. We can determine whether the phone is idle, randomly moving, or in cognitive/manual use.

Harald
So we’re sampling 20 to 100 Hz—really a lot of data—to train a model to determine whether the phone is idle, in motion, or being interacted with.

Katharina
Yes. Data efficiency is key. Even though the sampling is high, we process on the device and don’t send raw data. Important intelligence happens already on the device.

Harald
On our backend we receive like one data point per second. On the device it’s much higher, but we only transfer meaningful results.

Katharina
Exactly.

Harald
Another topic is crash detection. We call it AID, automatic impact detection. Briefly, as we’ll do a full episode later.

Katharina
AID is vital. If you’re in an accident, the emergency chain should start immediately. We listen for low-level signals from the smartphone’s accelerometer.

Harald
What does an accelerometer do? For non-technical listeners.

Katharina
It helps us understand directional movement of the smartphone. In a crash, you’d see peak acceleration followed by a drop. That’s our accident template.

Harald
Like when you turn your phone from portrait to landscape mode. We use the same sensors to detect impacts.

Katharina
Yes. We have labeled accident data and non-accident samples to train our model to detect impacts.

Harald
Great. So in simple terms, what is machine learning? How do machines learn?

Katharina
Machine learning is a collection of algorithms that learn from historical data to predict future outcomes.

With labeled data (supervised learning), we tell the machine this is class A or B.

Harald
So with smartphone usage, we tell the model: this motion was phone pickup, this was usage. The model learns from thousands of samples.

Katharina
Exactly. That’s supervised learning. But labeled data is expensive.

Unsupervised learning handles cases without labeled data using techniques like clustering.

Harald
So no human labeling, but we can cluster users based on behavioral patterns, assuming certain risk profiles.

Katharina
Yes, based on data and domain knowledge.

Harald
So three methodologies:

Supervised learning (human-labeled),

Unsupervised (cluster-based),

Heuristics (hard-coded rules).

Like if a car continues after one minute, probably no crash happened.

Katharina
Exactly.

Harald
Where will we use data science and machine learning in the future in telematics?

Katharina
Predictive modeling. Anticipating risk. For example, understanding when and where you’ll be to send weather alerts. Better to prevent than react to an accident.

Harald
Right. People have stable locations—home, work, shops, gym, etc.—so we use past behavior to predict future behavior.

Katharina
I 100% agree.

Harald
We’re also using generative AI for content creation—like personalized weather warnings. You provide the data science, we build on meteorological data and use generative AI to write messages in your language and tone.

Katharina
Exactly.

Harald
In conclusion, data science helps us turn raw data into meaningful information.

Katharina
And supports us making better decisions.

Harald
We differentiate data science as the human-aiding analytics, and AI as the machine-aiding automation. You framed it as: data science helps humans do the job, AI helps machines do it for us.

Katharina
Exactly.

Harald
Under the hood, AI powers large language models like ChatGPT. We’re aiming to predict risk and help people avoid accidents—prevention over reaction.

Katharina, it was a pleasure. Thanks for being here. And thanks to our listeners and viewers. Please subscribe and check the text below for more.