Advancing Dolphin Research Through AI Innovation
At a time where artificial intelligence (AI) seems to be pervasive in our everyday lives, NMMF scientists and veterinarians are putting it to work in ways that push the boundaries of marine mammal health and research.
We’re proud to share not one, not two, but THREE studies from our NMMF and Navy teams that highlight the recent results of machine learning directly supporting the care and conservation of dolphins at home and around the world. From the sound of a whistle to numbers on a blood panel, we’re walking alongside the AI innovation in the human health sector to enhance both personalized medicine for animals, and ultimately, understanding larger conservation needs.
Are you feeling ok? You don’t sound so great.
Most of us can probably remember a time when we came down with a slight cold, yet trudged into work. Perhaps we tried to mask the signs of our less-than-healthy state, greeting others in what we thought was a normal tone. But then, to your surprise, a concerned co-worker asks, “Hey, are you feeling ok?”
What gave you away? Maybe your voice was the telltale sign.
Vocal biomarkers are subtle changes or features in the voice that can give clues to the health status of the speaker. The human health sector has made incredible strides in identifying vocal biomarkers using machine learning models to diagnose conditions like depression, Parkinsons, and Covid-19. Now, a novel study suggests that vocal biomarkers in a dolphin’s signature whistle may be helpful in the early detection of abnormal health conditions.
Dolphins are known for being stoic animals, often showing signs of illness only after a condition has progressed very far. Being able to acoustically monitor a pod of dolphins and identify illness early could lead to early intervention for the animal and an overall better understanding of the health of the ocean. In this 2024 study titled, “Dolphin Health Classifications From Whistle Features”, NMMF and Navy scientists recorded over 36,000 signature whistles from 15 dolphins at the U.S. Navy Marine Mammal Program across five-years. To do so, the team conducted weekly “acoustic physical” exams, where dolphins were trained to produce their signature whistle in response to a hand signal (and after given a yummy fish reward). These whistles were then compared to the animal’s expansive health database kept by NMMF and Navy veterinarians to identify times of normal and abnormal health, including times of gastrointestinal issues, infections, and more “critical” diagnoses. Those whistles and their health labels were then used to train a machine learning model, with the hope that the computer could pick up on any subtle vocal biomarkers unrecognizable by human eyes or ears (see the spectrograms below to see just how similar dolphin whistles look across health states). Incredibly the model was 72% accurate at identifying a sick dolphin. Even more so, when looking at what were deemed “critical” health cases, the model was 94% accurate in its diagnosis.
Figure 1 from the Dolphin Health Classifications from Whistle Features study (Jones et al., 2024) shows just how similar dolphin whistles can visually look, even when produced in different health states. These spectrograms are how researchers can see the sounds they record, with frequency (in kilohertz) on the y-axis and time (in seconds) on the x-axis. The left-hand columns show two signature whistle examples for three of the participating dolphins who had a normal health state at the time of recording. On the right, whistles from the same dolphins recorded during various abnormal health states.
This is incredible news for dolphins in both professional care and in the wild, and an exciting step forward in using artificial intelligence for our conservation work. So, the next time you ponder about dolphins whistling beneath the waves, consider that these sounds might hold clues not just to who they are, but how they are doing. If we listen closely enough, we might also hear when the ocean itself needs care.
Reading between the (blood panel) lines
Blood tests are one of the most routine tools in medicine, both for people and for dolphins. A doctor draws blood, runs a panel, and compares the results to a standard “normal” range. But what if the most useful information isn’t just the numbers themselves, but the trend of the numbers at the patient level?
That’s the idea behind a new study titled “Machine Learning Discriminates Bacterial, Fungal, and Viral Infections Using Temporal Blood Analyte Dynamics in Bottlenose Dolphins”. When a dolphin gets sick, veterinarians often need to start treatment before they know exactly what’s causing the illness, both to manage symptoms and because results from more definitive tests like cultures can take days to come back. When treating infections, time is of the essence and an accurate diagnosis of a bacterial, fungal, or viral infection would be critical to avoid improper care and potential antimicrobial drug resistance. So the team asked, could a computer spot patterns in the biomarkers of a dolphin’s bloodwork that would reveal the source of the infection before test results can confirm it? And can pattern recognition at the individual level lead to earlier intervention?
Using 30 years of blood samples from dolphins with confirmed infections, the researchers trained a machine learning model to monitor changes in blood biomarkers, not just on single values from a single day, but on how each dolphin’s own values shifted over time compared to their personal healthy baseline. This is a bit like your smart-watch notifying you that your resting heart rate has been quietly climbing over the week and to be mindful of your activity levels, something a single reading could never say.
Incredibly the model correctly identified the type of infection around 76% of the time, with the highest accuracy in identifying fungal infections (80%) and nearly as well with viral (75%) and bacterial (73%) cases. The main biomarkers? Changes in eosinophil values (a type of white blood cell) and total White Blood Cell values that rose, fell, or stayed consistent depending on infection type compared to healthy levels.
Just as your own resting heart rate or sleep patterns can flag “something’s off” before you feel sick, this study shows that comparing a dolphin’s bloodwork against itself overtime, rather than checking it against a generic reference chart, can reveal a clearer story behind an illness. Sometimes, the clearest diagnosis doesn’t come from a chart of what’s “normal”, it comes from knowing what’s normal for you.
I’d know that voice anywhere!
Bottlenose dolphins have incredible vocal repertoires and complex pod structures that make identifying oneself essential. From maintaining group cohesion, to mother/calf bonding and socializing, dolphin clicks and whistles can hold important information about the speaker and their environment. Since their discovery in the 1960s, we know that dolphins produce a signature whistle, a whistle unique to a dolphin that forms within the first year of life and acts like a name tag. Researchers can use signature whistles in passive acoustic monitoring studies to identify the number of individuals in a pod and track pod movements. But dolphins can also share different whistle types within a pair or across a pod (like a group’s catchphrase), and can even mimic the signature whistles of other dolphins (like calling your friend’s name to get their attention).
How, then, can one dolphin know the identity of the whistler when it’s not making its signature whistle? And how can researchers get an accurate understanding of pod structure if dolphins are able to change the type of whistle they use?
Just like humans can identify the voices of their parents or friends, our most recent publication, “Identifying Dolphin Whistle Producers With Deep Learning: Moving Beyond Signature Whistles” sheds some light on what might be encoded in a dolphin whistle that makes one dolphin unique from others. Over 35,000 whistles from 20 different dolphins were used in the model, and the results were pretty remarkable.
The Sound and Health team performs an “acoustic physical” by asking the dolphin to produce its signature whistle on a “thumbs-up” hand signal. The computer screen on the left shows the live-feed of the ongoing recording showing this dolphin’s signature whistle as it appears.
When the model was trained and tested on every whistle type, it correctly identified the dolphin who made the whistle 85% of the time! Unsurprisingly, the model was really good at identifying the owner of a signature whistle, with over 96% accuracy. That alone is impressive, but here’s the really cool part: even when dolphins were sharing whistles, the model could still tell them apart, correctly identifying the producer of a shared whistle 93% of the time. That means there’s something about the way an individual dolphin whistles, maybe subtle differences in pitch, rhythm, or tone, that carries through no matter what whistle type it’s making. On the other hand, copied signature whistles of other dolphins proved to be the trickiest category (only 17% accuracy). So, while you might think you can do a really good impression of your brother, a computer model might think otherwise. However, the researchers only had a small number of examples to work with in this category, so this is a puzzle still waiting to be solved with more data.
So, what does this mean for our ocean neighbors?
Understanding that individuality can shine through even when dolphins aren’t using their “name tag” whistle opens up exciting new possibilities for eavesdropping on dolphin conversations. Tools like this could help scientists figure out not just how many dolphins are nearby, but exactly who is present and who’s doing the talking. Turns out, whether it’s a human hello or a dolphin whistle, a voice can say a lot more than just the words.
