A study can contain 602 videos and still represent only 26 people. That is the first number to keep in view when reading the results from CiliaTracks, an open computational framework that analyses how fluorescent particles move across cultured airway cells. Its models classified videos from primary ciliary dyskinesia samples with 95–97% accuracy. The result is promising because it makes ciliary transport measurable. It is preliminary because many recordings came from the same donors.
Primary ciliary dyskinesia, or PCD, is a genetic disorder in which motile cilia do not move mucus normally. Diagnosis is difficult because the disease can arise from different genes and produce varied motion patterns. The researchers cultured airway epithelial cells, placed fluorescent microspheres on the moving cilia and recorded the particles with high-speed microscopy. The tracks become a proxy for how effectively the cilia transport material.[1][2]
The denominator behind the accuracy
| Measure | Value | Context |
|---|---|---|
| People with genetically confirmed PCD | 14 | Their cells contributed 301 PCD-labelled videos.Source 1Source 2 |
| Healthy donors | 10 | Part of the non-PCD comparison group.Source 1Source 2 |
| People with cystic fibrosis | 2 | Included to test whether abnormal respiratory disease alone resembled PCD.Source 1Source 2 |
| Total videos | 602 | 301 PCD and 301 non-PCD.Source 1Source 2 |
| Reported model accuracy | 95–97% | Across classical feature-based models and a convolutional neural network.Source 1Source 2 |
The model watches the cargo, not the cilia directly
Cilia beat too quickly and intricately for a single visual impression to describe all of their function. By following beads moved by the cilia, CiliaTracks measures the transport outcome. The software detects particles, links their positions across frames and extracts features such as speed, direction, displacement and the shape of the trajectory.
The biological and computational pipeline
| Stage | What happens | What the output means |
|---|---|---|
| Cell culture | Airway epithelial cells are grown until they form a multiciliated surface | A laboratory model of the donor’s ciliary system.Source 1Source 2 |
| High-speed video | Fluorescent microspheres move across the beating cilia | A recording of cilia-driven transport rather than a direct patient symptom.Source 1Source 2 |
| Tracking | Software follows each bead and extracts kinematic features | A numerical description of direction, speed and organisation.Source 1Source 2 |
| Feature model | Classical machine learning uses engineered transport measurements | A classification based on interpretable trajectory statistics.Source 1Source 2 |
| Image model | A convolutional network learns from trajectory images | A classification from spatial patterns without relying only on named features.Source 1Source 2 |
Why the video count can exaggerate confidence
If one donor contributes many videos, those recordings share genetics, cell culture conditions and laboratory handling. A model can appear highly accurate by learning donor-specific or batch-specific patterns that recur across that person’s videos. The key validation question is therefore not only whether held-out videos were classified correctly, but whether every recording from a donor was kept together when training and testing were separated.
The paper reports strong discrimination and the ability to distinguish PCD from normal cells or cystic fibrosis. The next evidence should include larger donor cohorts, explicit participant-level splitting, independent laboratories and prospective comparison with the full diagnostic pathway. This is an inference from the study design, not a claim that the authors mixed videos from the same donor across every split.
Objectivity is useful even before diagnosis is automated
PCD assessment can include genetics, nasal nitric oxide, microscopy and other specialist tests. CiliaTracks need not replace that pathway to be valuable. A quantitative transport signature could help laboratories compare samples, identify unusual phenotypes and reduce disagreement caused by subjective visual interpretation.
The two analytical routes are also useful checks on one another. If a feature model and an image model succeed for different reasons, researchers can ask which transport properties are stable across genes and laboratories. If performance collapses outside the original site, that result would reveal which parts of the signature were local artefacts rather than disease biology.
The study’s strongest contribution is a new measurement object: the path of cargo moved by living cilia. The honest reading keeps both denominators visible. There are 602 recordings, which makes the computational experiment substantial; there are 26 people, which makes clinical generalisation a much larger task still.
Sources & further reading
- Machine learning analysis of cilia-driven particle transport distinguishes primary ciliary dyskinesia cilia from normal cilia Communications Biology · 12 Oct 2026 · Accessed 16 Oct 2026
- Machine Learning Analysis of Cilia-Driven Particle Transport Distinguishes Primary Ciliary Dyskinesia Cilia from Normal Cilia PubMed Central · 4 Nov 2025 · Accessed 16 Oct 2026