Two days before New Year’s Eve, a neonatal team called in distress. Nurses refused to work in newly built single-room NICU suites because they could no longer hear the babies. In that moment, a familiar sound—crying—had vanished. What followed became one of the clearest examples of hospital-driven innovation in our history.
Single-room neonatal care promised calmer environments and stronger bonding between parents and their premature infants. But as one hospital learned, silence can also become unsafe.
The incubators in these new suites were fitted with acoustic hoods. The walls were insulated to protect privacy. The background machinery created a soft but constant hum. And together, these advances removed one of the earliest danger signals in neonatal care: the sound of a baby crying. The nurses raised the alarm. They simply could not work safely if they could no longer hear their patients.
This was the moment the hospital reached out to Neolook - just before New Year’s Eve - asking whether we could restore what architecture had taken away: situational awareness.
The request quickly turned into a head-to-head competition between vendors. The hospital needed a fully functioning demonstrator that could operate in a real NICU environment, withstand 24/7 clinical use, and integrate with existing systems. The competition window was tight. The clinical stakes were high. And the environment was more challenging than it seemed.
A neonatal unit contains a constant 50 dB background noise from equipment. Nurses talk. Parents visit. Alarms sound. Microphones must be placed outside the incubator, where the baby’s cry is often quieter than everything around it. Detecting a premature infant’s cry under these conditions wouldn’t just require audio sensing. It would require a serious AI system.
Premature infants do cry—but less often than healthy term babies, and often with lower intensity. Capturing cry samples at the right moment was itself a challenge. In the early days of the project, data was the real bottleneck. Cry events were unpredictable. Twins created overlapping sound patterns. Rooms varied in acoustics. And large periods consisted simply of silence. A high-performance algorithm needed tens of thousands of real-world samples. But the NICU could only generate them slowly.
Collecting enough input required months of recording, annotation, and refinement. Only once we passed 100,000 fragments could the model mature to the reliability the hospital needed.
Detecting crying by volume alone is impossible in the NICU. Instead, the system “sees” crying. Every audio fragment is converted into a spectrogram—an image representing frequencies over time. The system recognises patterns in these images, not in the raw sound. A convolutional neural network learned to distinguish true neonatal cries from everything else: pumps, footsteps, conversations, monitor tones, accidental noises.
Early training was skewed used with several thousands samples annotated by hand. Over time, the dataset expanded dramatically. With each wave of new clinical recordings, the system improved, becoming more selective and more robust.
The model achieved 90% accuracy - but accuracy was not the real goal. Reliability was NICU nurses cannot afford false alarms, and they cannot afford missed cries. So the system was tuned to favour confident, meaningful alerts.
From the beginning, the cry detection engine was integrated into our platform architecture.
It works alongside:
This flexibility turned out to be essential. Technology does not replace clinical judgment - it supports it.
The system was also built with strong privacy boundaries. Audio fragments are processed in real time and can be configured for extremely short retention windows. No long-term audio archives are required for operation. Data that must be retained for quality monitoring follows strict retention policies.
The algorithm prioritises selectiveness to prevent alarm fatigue. And the configuration panel supports local clinical policy, ensuring each hospital can set thresholds according to its own protocols and safety requirements.
When the hospital evaluated the demonstrators, four things stood out.
Neolook’s system worked reliably in real-world sound conditions.
These factors allowed Neolook to deploy the system in the full departement opertionally and win a new contract.
The silent rooms now had a voice again.
Cry detection began as a one-hospital problem. Today it sits inside a wider neonatal AI ecosystem.
Each installation generates new anonymised training fragments, strengthening the model. Each upgrade increases selectiveness. Each new NICU adds diversity to the dataset—different room acoustics, different incubator models, different patient groups. This is the power of hospital-driven innovation: the real world becomes the engine of progress.
Moving on in time Dutch NICUs adopt the updated cry detection module, and Neolook integrated it with S2S Family, S2S Professional on a single all-in-one platform.
What started as a moment of silence two days before New Year’s Eve has now helped shape a new generation of neonatal safety technology—one grounded in need, refined by real-world data, and amplified by the collective experience of hospitals across Europe.
Single-room NICU care continues to expand globally. And with it comes the same question: How do we give nurses presence when the architecture removes their senses? Cry detection became our first answer. Multimodal AI will deliver the rest.
But this project taught us something important: innovation begins where the environment gives clinicians room to innocate if the need becomes undeniable.
The hospital that reached out in silence helped set in motion a new era of neonatal intelligence - one powered by the voices of the smallest patients, and by the professionals determined to hear them.