Ana Laguna is a BBVA Data Scientist. And also mother. Now she is going to be so in a couple of months, and since her first pregnancy she had the idea of taking advantage of her knowledge in artificial intelligence (AI) to solve her own fears and answer a crucial question for her at that moment: why is she crying? my baby? Thus, he has been able to develop an algorithm that listens to the cry of the little one and identifies what is the cause that motivates the cry. Today his goal is to take advantage of this project to create an application that is capable of diagnosing diseases in babies by their type of cry.
How did Ana come up with such an idea? “This is a project that I carried out during my first baby’s maternity leave. She must have been very bored, and as a hobby I started doing it. Now, as I am going to have another one, I have taken up the subject and I am going to take it more seriously to see what comes out of here ”.
Lest we think that Ana is a crazy mother who experiments with her children, she puts us a bit in context so that we understand her. During pregnancy, everything is very beautiful, but there are also contradictory feelings such as fear, uncertainty, doubts about whether we will be good parents and how they will react when the baby arrives. Will we understand it? Will we get their education right? “And this is only”, says Ana Laguna, “in the process of pregnancy, when the baby is not yet here.”
But the situation is complicated when she gives birth, because mixed feelings appear and postpartum depression, which affects 60% of women due to hormonal changes, and which can make the mother reject the baby. When we go home with our baby we do not have any technological device to instruct us on what to do with the baby. “At most we have our grandmother and our mother. And then the baby starts crying, and we don’t know what’s wrong with him. If the pregnancy has been a master’s degree, now I’m going for a doctorate ”, explains the BBVA Data scientist.
In this context, and even being in her first pregnancy, Ana wonders if there is an Asian application that helps to manage this problem. He started looking, just two years ago, and discovered that there was a Chinese app, with very bad comments. Even so, for reasons of prosody, which is a branch of linguistics that studies the accent and intonation of a language, what is heard from the baby in the womb if the mother is Chinese or Spanish does not mean the same thing, and the crying is also different. For these reasons, the Chinese app did not work for Ana.
A lot of data was needed
This is how this expert in data science and artificial intelligence thinks that when her first baby is born, she will try to do her own study. The first challenge in the project is that it needed data, a lot of data. And he starts looking for online crying databases, but he can’t find them. “And it was not a question of going to a hospital in the newborn section to start recording crying.”
So from the time her first baby is one month old, and until the fourth month of life, when she cried, and at least once a day, she tries to record it to find out what was happening to her. In total, Ana recorded about 70 audios. At first he wanted to recognize from when a mosquito bit him to when a pimple came out. But then she realized that the baby’s life cycle was pretty routine and that it’s the same all the time. The most frequent ‘baby modes’ of the baby were sleeping, hunger, gas and diaper changes. And then there were others like he is standing up, has infant colic pains, or needed pampering. And this is how Ana arrives at the seven labels that her algorithm could begin to learn.
The second challenge was to tag all the audios, and the third the versatility. Ana Laguna says that her baby was a glutton and cried a lot because he was hungry. “And that is why I find a data share that tells me that 50% of the cries are due to hunger, 3% from pampering, another 3% from sleep and another 3% from anger. I realize that I don’t have versatility and I have to go from a multi-classification algorithm to a binary classification algorithm hungry or not hungry ”.
Crying patterns, if they could be heard, are very different. The hunger, with the mime, for example, are very different. And this is what the human ear receives. The algorithm receives spectrograms, which is a graphical representation of the audio signal in time space. And this is how the algorithm seems to be that it also detects the patterns, through these spectrograms.
Ana’s first conclusions in this experiment with her first baby were that she had models with good precision, on the other hand, the data sample was very small for this type of algorithm, and also had a significant bias because she only had data from her baby.
Baby Crying Donation Campaign
Now, as the new baby is on the way, and the social interest that seems to have aroused her project, since the beginning of the year, through the NGO So Good Data, Ana has launched a baby crying donation campaign in which she requests the mother’s label, the father’s and mother’s nationality for the prosody reasons discussed above, the baby’s months of life, between 0 and six months, and gender.
With this campaign, Ana has achieved in just over a month some 300 audios, and applying future engineering AI techniques and data augmentation has created a data share of 1200 audios, which no longer have anything to do with the 70 with which she started. . And at the level of versatility of the data, it already has much more variability: 24% hunger, 23% pampering, and 21% sleep. And it also has much more data from other countries.
“The current situation with all the recovered data”, affirms Ana Laguna, “is that I have no bias, although the precision of the model, which is what tells me if I am doing it well or not, is still 70%, and I have to improve it ”.
The next steps for the BBVA scientist is to get more mothers involved, more audios, balance the data share , get mothers of other nationalities, use other deep learning techniques, in addition to transfer learning , and try tagging with clustering issues, for example. Their goal is to optimize the algorithm and improve accuracy to get an app out as soon as possible so that it can help mothers.
“I view this project in the context of Data for Social Goods because I think it can have a social impact and be very useful for new parents. Probably a very experienced mother will tell me that you do not need artificial intelligence to know what your baby needs. They are the ethical questions that always come up in these types of advances. But in that first month of the baby’s life that I didn’t know what was wrong with him, it would have been useful to have an app that would help me recognize why he was crying ”.
Ana would love to take this project to the early diagnosis of diseases, such as autism, thanks to the fact that artificial intelligence can identify it from crying. If you want to help Ana, send her your baby’s cry. You just have to do it through the NGO he has founded, So Good Data .
