Systematic review of artificial intelligence and digital-automated tools for screening, diagnosis and characterisation of primary ciliary dyskinesia

Abstract

Background

Primary ciliary dyskinesia (PCD) is a rare, multigenic disorder of impaired mucociliary clearance leading to a spectrum of disease including chronic respiratory infection and bronchiectasis. PCD is phenotypically variable causing diagnostic challenges. PCD is underdiagnosed, and when it is made, diagnosis is often delayed. We describe evidence for use of digital-automated applications and/or artificial intelligence (AI) to improve PCD screening, diagnosis and characterisation.

Methods

A systematic literature search (2004–2025) was conducted across PubMed, Embase and BioRxiv/MedRxiv (PROSPERO:CRD42024605689) by members of the ERS BEAT-PCD Clinical Research Collaboration (CRC). Screening and data extraction was performed by two independent reviewers and findings summarised qualitatively. In collaboration with the digital health ERS CONNECT CRC network, a narrative discussion focused on steps to real-world implementation.

Results

Of 770 screened abstracts, 73 full-texts were assessed and 28 PCD-relevant studies with automated digital and/or AI-driven tools were included in the qualitative synthesis. Tools for enhanced ciliary function, ciliary ultrastructure and ciliary protein immunofluorescence analysis were presented in 17 studies. 11 papers presented tools to screen e-health records for PCD, stratify patients by genotype and phenotype, or characterise computed tomography parameters.

Conclusions

Tools are available that could improve PCD diagnostic accuracy and reduce time-to-diagnosis. Studies were often single-centre, retrospective and of small sample size. The development of automated-digital tools or AI-driven tools requires representative patient datasets, human expert judgement to interpret and continual safety auditing. External validation is paramount before adaptation of healthcare systems, with adherence to EU AI Act framework to ensure accountability and safeguard against misuse. Simplified tools and philanthropic partnerships could facilitate implementation of new systems in resource limited settings.

Shareable abstract

Artificial intelligence and digital-automated tools could improve diagnostic accuracy in PCD and reduce time to diagnosis. However, available evidence is largely restricted to single-centre, retrospective and small sample size studies. https://bit.ly/48LPRAc

Introduction

Artificial intelligence (AI) and digital automation are increasingly being applied in healthcare to clinical workflows where interpretation of data is complex, time-consuming or dependent on specialist expertise. Automated digital tools, defined as deterministic, rule-based approaches that standardise or quantify measurements (for example, video and image processing pipelines), and AI-driven tools, which use learning-based methods (including machine learning and deep learning) can be used to classify, predict or extract patterns from imaging and clinical datasets [1–3]. Overall, application of these tools in the healthcare setting may result in expedited workflows, increased healthcare system capacity and improved patient satisfaction [2, 3]. In the case of rare diseases, they also have the capacity to optimise clinical trials as well as streamline and expedite diagnosis and treatment management despite limited patient numbers [4].

Primary ciliary dyskinesia (PCD) is a rare inherited multigenic group of motile cilia disorders causing mucociliary clearance dysfunction and progressive airways disease with an approximate incidence of 1:7500 globally [5]. PCD is currently underdiagnosed due to lack of awareness amongst healthcare professionals and significant expertise and resources required to conduct diagnostic tests. A diagnosis is confirmed by ciliary ultrastructural defects confirmed by transmission electron microscopy (TEM) [6] and/or pathogenic variant(s) identified in any of >55 causative genes [7] (supplementary figure S1). ∼30% of people with PCD lack a TEM and/or genetic diagnosis. Adjunct investigations support a highly likely diagnosis of PCD instead [8]. Nasal nitric oxide (nNO) gas measurements are typically below 77 nL·min⁻¹ in PCD; high-speed video microscopy (HSVM) is used to assess ciliary function abnormalities but is confounded by secondary dyskinesia, and air–liquid interface culture methods are employed to improve patient sample health to repeat analysis; and immunofluorescence microscopy analysis can demonstrate absence or mislocalisation of ciliary proteins [8]. The genotype and diagnostic phenotype relationship is summarised in supplementary figure S1.

The use of automated digital and AI-driven tools in PCD has the potential to support standardisation, reproducibility and efficiency within specialist diagnostic pathways and respiratory research [9, 10]. The European Respiratory Society (ERS) BEAT-PCD Clinical Research Collaboration (CRC) is a Pan-European network bringing together clinicians, scientists and patients to improve the diagnosis, management and treatment of PCD [11]. In line with its mission to advance diagnostic and clinical innovation in PCD, we undertook this systematic review within the BEAT-PCD framework.

We aimed to summarise, in a qualitative synthesis, the available evidence supporting the use of automated digital tools or AI-driven tools in the screening, diagnosis and characterisation of PCD. Additionally, through a focus group narrative, developed with ERS CONNECT CRC network [12], we aimed to describe implementation considerations and future priorities for integrating these tools into PCD clinical care and research.

Methods

Search strategy

A systematic literature search was conducted from January 2004 to October 2025 across PubMed, Embase and BioRxiv/MedRxiv using algorithms consisting of combinations of keywords and medical subject headings relevant to motile cilia or PCD and automated, digital, computerised or AI tools. This timeframe was chosen to capture the evolution of automated digital and AI-driven methods in PCD, from early image-processing tools in the mid-2000s to contemporary AI-driven applications. Preliminary exploratory searches conducted by the authors during protocol development did not identify any studies applying AI-based or automated digital methodologies (excluding manuscripts on ciliary beat frequency measurements only) for PCD prior to 2004. Accordingly, 2004 was defined as the start date for the systematic search. Algorithms were adjusted as appropriate for each electronic database (supplementary table S1). Titles and abstracts of retrieved records were screened by four independent reviewers (P. Kouis, M. Bottier, N. Rumman and M. Shanks), ensuring that each record was screened by at least two independent reviewers. Disagreements were resolved independently by consensus between two senior researchers (A. Shoemark and C. Jackson). The references of full texts were also screened for additional eligible studies. The review adhered to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines and was registered to International Prospective Register of Systematic Reviews (PROSPERO ID: CRD42024605689).

Study selection and data extraction

Retrieved records and pre-prints were considered eligible for inclusion if they described the implementation or development of automated digital and AI-driven tools aimed at: 1) improving the accuracy of established diagnostic PCD tests (including nNO, HSVM, TEM, immunofluorescence, genetics) or exploring additional discriminative tools (Theme 1); or 2) improving PCD screening and disease characterisation, including phenotype–genotype correlations in PCD (Theme 2). Studies relating to automated digital tools for ciliary beat frequency (CBF) assessment were excluded, due to superseding studies that incorporate ciliary beat pattern (CBP) analysis for PCD diagnostics [13]. Narrative reviews, opinion pieces and editorials were excluded. Given the recent developments in automated digital and AI-driven tools, abstracts published in conference proceedings were not excluded a priori.

Data variables were extracted independently by P. Kouis, M. Bottier, N. Rumman, M. Shanks, A. Protopapas and C. Jackson onto a standard form: Author; Year; Study type; Population (Suspect PCD, confirmed PCD, hospital cohort); Application (Diagnosis/screening/disease characterisation); Research question (aim); Data source (Dataset); Reference test/method; Automated digital tool – AI-driven tool description; Outcomes quantified; Findings/conclusions; and Limitations and challenges.

Risk of bias assessment

The methodological quality of included studies was assessed using modified versions of the CLARITY Group at McMaster University Tools to Assess Risk of Bias in Case–Control Studies and Cohort Studies [14]. The CLARITY tool includes five key domains for case–control studies and eight domains for cohort studies, each addressing a specific aspect of study design and execution. Both tools were modified (supplementary table S2) to better reflect methodological considerations relevant to the present review. Each domain was rated on a scale from “Definitely yes” (indicating low risk of bias) to “Definitely no” (indicating high risk of bias), with intermediate options of “Probably yes” and “Probably no”.

Qualitative synthesis of published evidence

Due to the expected heterogeneity in study designs and outcomes of included records, a quantitative synthesis was not planned a priori, and findings were summarised qualitatively. The main study characteristics were summarised in table format with particular emphasis on the specific aims, and methods, including the description of the AI or automated digital tool, as well as key findings and main limitations or challenges. Furthermore, the narrative qualitative synthesis includes a brief description of the aim and method employed in each study and a short discussion of the pros and cons of each application in PCD.

Focus group narrative: implementation considerations and future priorities

To address implementation considerations and future priorities for integrating automated digital and AI-driven tools into PCD care, we conducted a structured expert focus group discussion. ERS CONNECT CRC members J. Buekers and D. Megaritis were provided with the manuscript, tables and figures 2 weeks prior to a focus group discussion led by C. Jackson and M. Bottier. Two questions were raised, structured around validity of automated digital and AI-driven tools, planning and implementation of pipelines, regulatory and ethical considerations and limited resource settings:

1. What are the practical steps and ethical considerations needed for integrative use of automatic digital and AI-driven tools to improve patient screening and PCD diagnostic accuracy, and to better stratify diagnosed patients for tailored care and research?

2. What methods/examples/considerations will support resource-limited areas to include use of automatic digital and AI-driven tools despite patient clinical/diagnostic data gaps?

The transcript was automatically captured via Microsoft Teams and summarised by question and tabulated by C. Jackson and finally agreed by all focus group members.

Results

Of 770 systematically retrieved records, 73 studies were eligible for full-text assessment, and 28 were ultimately included. 15 included records described AI-driven tools. The PRISMA diagram summarising the systematic review is shown in figure 1.

FIGURE 1.

PRISMA flowchart for literature screening and inclusion. AI: artificial intelligence; CBF: ciliary beat frequency.

Study characteristics

Most studies were single-centre and retrospective. Sample size varied and occasionally was either not disclosed or only one patient, also several studies included data from only suspected PCD cases, and not all records demonstrated test outcomes that would confirm PCD using the current guideline [8] (table 1).

TABLE 1.

Theme 1: characteristics of studies focusing on digital-automated tools or artificial intelligence (AI)-driven tools used to enhance the analysis of ciliary function, ultrastructure or immunofluorescence in primary ciliary dyskinesia (PCD)

First author, year [ref.]	Study population	Diagnostic application	Research question/aim of study	Digital/automated or AI method employed	Reference test/method	Main findings	Limitations/challenges
Application in ciliary function analysis
Renaud, 2024 [19]	Primary diffuse chronic rhinosinusitis (CRS) patients (n=41) classified by a European Position Paper 2020 and healthy controls (n=43)	HSVM	Evaluate the ability of ciliary beating efficiency (CBE) to discriminate CRS patients from controls	Computational flow model with microbead tracking for estimation of CBE Index	n/a	CBE index was significantly lower in the CRS group than in the controls CBE index diagnostic accuracy for CBE equal to 0.34 mPa: o sensitivity: 94%, specificity: 76% o AUC: 0.90	Large cohorts required to address CBE and diagnostic accuracy considering different types of primary diffuse CRS
Schneiter, 2023 [20]	Confirmed PCD patients (n=4) and healthy controls (n=7)	HSVM	Present open-source HSVM image analysis software to extract quantitative information about ciliary motility and mucociliary transport	Cilialyser software: Carries out image stabilisation using Python Cilialyser estimates the HSVM video activity map and the frequency correlation (mean distance over which the CBF remains preserved) Particle tracking module relies on tracers extracted with Lee-filtering followed by Gaussian filtering	Genetics	Cilialyser can improve and facilitate the analysis of ciliary function captured by HSVM: Image stabilisation can improve measurement of CBF In healthy samples, the maximum transport speed ranged from 15.4 to 81.1 μm·s−1. In all samples derived from the PCD patients, all measured transport speed values were lower than 0.4 μm·s−1 Median frequency correlation length between 20 and 30 μm in healthy controls versus 0 in PCD	Does not classify normal/abnormal phenotypes Requires user input, expertise and judgement Diagnostic utility of extracted quantified parameters remains unclear Further validation required in cohorts of PCD suspect patients
Chioccioli, 2019 [22]#	Confirmed PCD patients (n=8) and healthy controls (n=3)	HSVM	Determine a set of quantitative measures using a computer interface for manually tracking the cilium position over time creating a continuous profile of the cilia waveform. Assess the discriminative ability of these measures	The quantitative HSVM analysis strategy led to the estimation of the following quantitative parameters: Amplitude of the beating stroke Standard deviation of the curvature of the cilium (at mid-length) Mean force the cilium exerted on the fluid (using resistive force theory) Principal component analysis and unsupervised k-means clustering algorithm used to test if these parameters distinguish between healthy controls and samples from PCD patients	Genetics	Principal component analysis on the set estimated parameters (“barcode”) allowed for discrimination between healthy and two variants of PCD	There were a small number of samples and only two genetic variants assessed Additional studies using larger datasets with wide spectrum of genetic variants are needed to validate these findings
Bottier, 2017 [18]	Suspected PCD patients (n=11)	HSVM	Assess the use of microbeads tracking to characterise mucociliary flow and estimate ciliary beating efficiency (CBE)	Computational flow model with microbead tracking	HSVM and TEM	CBE assessed using microbeads can be used as a PCD diagnostic index Reflects the shear stress exerted by the cilia wall on the fluid CBE ranged from 0.6 to 9.5 mPa	No follow-up experiments to test the diagnostic accuracy of this parameter with a PCD/non-PCD patients’ cohort
Quinn, 2015 [17]#	1st cohort: healthy controls (n=27), confirmed PCD patients (n=5), CHD with abnormal ciliary motility (n=17) 2nd cohort: heterotaxy patients (n=27; with abnormal ciliary motility n= 10; with normal ciliary motility n=17), confirmed PCD patients (n=4)	HSVM	Present a computational pipeline to decompose ciliary motion into quantified components and assess classification of ciliary motion as normal or abnormal	Applied optical flow analysis using spatial and temporal derivatives to quantify ciliary motion vectors across video frames Extracted instantaneous rotation and shear deformation parameters within regions of interest Used autoregressive modelling with PCA to capture periodic motion variance	Manual beat pattern determination by blinded experts	Software could be used for automated or manual assignment of normal versus abnormal HSVM phenotype: Rotation and deformation computed for healthy cilia showed strong periodic behaviour and large magnitudes. In contrast, the rotation and deformation in abnormal cilia showed erratic, weak periodic behaviour in addition to reduced magnitudes Accuracy was over 70% for parameters and fairly agreed across both datasets. In contrast, CBF accuracy was only 50%	Expertise required to select best video quality Software would not interpret poor quality sample or HSVM Gold standard (ground truth) established only based on HSVM interpretation by experts. Other diagnostic tests not described This method had high discrimination capacity, but it is not clear if mild CBP cases were included in the sample
Pifferi, 2013 [21]#	Suspected PCD patients (n=151)	HSVM	Evaluate if ciliary activity in suspension culture (spheroids) supported by computing analysis could by itself lead to a firm diagnosis of PCD, without the need for longer culture (air–liquid interface)	Soft computing/Artificial Neural Network (ANN)	HSVM and TEM	Suspension culture and ANN discriminate PCD from SCD: After 5 days of culture (faster than ALI culture) Accuracy of suspension spheroid culture parameters: o rotation (Sens: 0.79, Spec: 0.97) o migration (Sens: 0.17, Spec: 1.00) o remove debris (Sens: 0.07, Spec:1.00) o normal CBP (Sens:0.64, Spec: 1.00) Abnormal CBP (Sens: 0.33, Spec: 1.00)	TEM was not possible after spheroid culture Inclusion of atypical PCD cases in the “SCD” control group to train the ANN model cannot be excluded
Sears, 2013 [16]	Cultured HBE (n=2) and HNE (n=1) cells from healthy control subjects	HSVM	Develop an empirical system that allows a “smooth” representation of cilia movement dynamics that can be used to validate theoretical models of the ciliary beat	Quantitative motion modelling relying on Gaussian functions to define cilia curvatures, based on HSVM data perfectly orientated via manipulation of cultured samples	n/a	Main findings included: A framework to extract cilia waveforms A standard (normal) cilia waveform to be used for comparisons with altered waveforms Measures of cilia waveform variability	Model not applied or validated in samples from PCD patients Utility for PCD diagnosis was discussed as a possible application
Papon, 2012 [15]	Confirmed PCD patients (n=10), suspected PCD patients (n=9) and non-PCD (n=15)	HSVM	Assess how quantitative ciliary beating parameters were compared with qualitative evaluation in patients in whom the diagnosis of PCD was confirmed, excluded or was inconclusive	Quantitative analysis of HSVM recording to estimate the following ciliary beating parameters: Global frequency Power stroke duration Recovery duration Pauses after the active and recovery strokes Total pause Cilia length Beating angle Distance travelled by the cilium per second (i.e. the length of the path travelled by the tip in 1 s) Area swept per second (i.e. the area swept by the cilium in one second) Last two parameters were then weighted by the percentage of beating ciliated edges	Ciliary ultrastructural abnormalities on TEM combined with low nNO	Quantitative ciliary motility assessment is potentially useful to define dyskinesia and dyskinetic pattern for the diagnosis of PCD, especially in cases with mild CBP abnormalities: The AUC of the ROC curve was ≥0.69 for each parameter examined The distance travelled per second (by the cilium tip) and the area swept per second (by the cilium), weighted by the percentage of beating ciliated edges, presented the best AUC (0.984 and 0.977, respectively) Weighted distance travelled per second was 96% sensitive, 95% specific at cut-off of 24 μm Weighted area swept per second, a cut-off of 62 μm2 gave a sensitivity and specificity of 94% and 94%, respectively	TEM analysis used for diagnosis was pre-BEAT-PCD international consensus; therefore misclassification of some cases cannot be excluded. Based on clinical characteristics in the non-PCD groups, presence of some missed PCD cases is possible No genetics included in the gold standard Larger studies with more diagnostic certainty are needed to replicate findings
Application in ciliary ultrastructure analysis using TEM
Bottier, 2023 [29]#	Suspected PCD patients (n=158)	TEM	Assess whether computer vision and machine learning can aid in the diagnosis of PCD	Artificial intelligence platform (Intel Geti) was used to train a machine learning with 21 000 TEM ciliary cross-sections and then classify them based on normal or abnormal ultrastructure or “unusable” for diagnostic purposes (tilted or distorted images)	UK standard diagnostic pathway	Using retrospective and prospective patient samples we have found PCD-AID can reliably identify ciliary ultrastructural defects Sensitivity: 0.87; Specificity: 0.88 PCD-AID could assess TEM images in under 1 min per patient Good agreement with diagnostic specialists (>75%) Outperforms specialists at identifying subtle central pair defects associated with HYDIN variants	Combination of retrospective and prospective PCD suspect patients Larger cohorts required to validate these findings
Staar, 2023 [27]	Suspected PCD patients with inconclusive genetics (n=16)	TEM	Assess feasibility of the PCD Detect software for evaluating ciliary ultrastructure with regard to the confirmation of outer and inner dynein arm defects	Image averaging of ciliary cross-sections	TEM (hallmark defect)	Findings confirm that PCD Detect is an easy-to-apply and useful tool. Among suspected PCD with inconclusive genetics: Definite diagnosis (hallmark ultrastructural defect) was established in 6 out of 16 (37.5%) Identified central complex (class 2) defect in another patient, providing supporting evidence for genetic variants in NME5 as likely pathogenic for PCD	Subject to the quality of brushing Multiple, manually chosen and layered cutouts Investigator operating PCD Detect not blinded to genetic results
Felšöová, 2021 [30]#	Training set composed of cilia cross-sections: 1500 class 0 (normal) 400 class 1 (primary defects) 400 class 2 (secondary defects) Validation dataset of suspected PCD patients (10 random patients)	TEM	AI automated analysis of TEM images (PCD Quant software) to quantify class 1, 2 and secondary defects and mutual orientation of cilia	Automatic quantitative TEM analysis performed by a deep learning neural network model	Manual TEM analysis	Main findings: Compared to manual evaluation the PCD test PCD Quant diagnostic accuracy: o Sensitivity: 67.9% o Specificity: 87.9% Cohen's kappa (excluding background class misclassifications=0.718) PCD Quant was able to automatically analyse central complex defects and microtubular disorganisation and differentiate them from secondary defects Analysis time: up to a minute per image	Low sensitivity especially cilia cross-section with class I defects Validation with a larger dataset of suspect PCD patients is required Cannot be used as a stand-alone analysis tool because it fails to assess class 1 hallmark PCD defects Statistical errors due to image “noise” meaning microtubules and background are not differentiated
Shoemark, 2020 [26]#	Confirmed PCD patients (n=11) and healthy controls (n=3)	TEM	Provide and validate a tool to classify images of cilia cross-sections for subtle changes	Tool kit for image averaging of ciliary cross-sections and classification (Matlab) using principal component analysis hierarchical classification	Genetics	Enhancement of diagnosis for both straightforward and subtle, easily missed structural changes in TEM images: Averaged microtubular doublets show complete or sporadic absence of the ODAs in DNAH5 variants Multiple microtubular doublets classes generated from a subject with DNAH9 variants due to loss of the ODA at the distal cilium region Absence of IDAs in CCDC40 and CCDC103 variants Confirmation of HYDIN pathogenic variants by visualisation of the CC component C2b	Requires other tests to confirm diagnosis as not a gold standard technique Require multiple TEM images of well-orientated cilia and knowledge of how to interpret the final averaged images
Funkhouser, 2014 [25]	Suspected PCD patients (n=26)	TEM	Assess diagnostic test performance of a semi-automated TEM image analysis tool	Image alignment and averaging of ciliary cross-sections	Diagnosis based on clinical presentation, history, imaging, TEM, ciliary video microscopy, nNO and genetic sequencing	Improved inter-rater reliability Similar or better sensitivity and specificity compared to raw TEM images (across two observers)	Evaluates the performance of this method only for dynein arm defects and not for other PCD-related TEM abnormalities Results available only for two observers
Funkhouser, 2011 [24]	Suspected PCD patients (n=32)	TEM	Determine whether inter-rater reliability of PCD diagnosis could be improved via a computer image analysis algorithm for TEM images	Computer image analysis algorithm in MATLAB(c) to align axonemal microtubular pairs	Manual TEM analysis	Image alignment algorithm improved inter-rater reliability for evaluation of dynein arm loss in ciliary axonemes. TEM “gold-standard technique”: k=0.68 Using 10 computer-generated axonemal means per case: k=0.81 Using single computer-generated mean of all microtubular pairs per case: k=0.94	Unclear diagnostic status of PCD suspect cohort Doesn't assess MTD or CC defects Number of cross-sections assessed low based on BEAT-PCD international TEM consensus
O'Toole, 2012 [23]	· Samples (human airway cells) from a confirmed PCD patient, a healthy control and a CF control patient: 45 cilia cross-sections of CC complexes from Chlamydomonas  reinhardtii 46 cilia cross-sections of CC complexes from human airway cells	TEM	Compare the axonemal structure of human cilia and C. reinhardtii flagella to identify similarities and differences Assess how the combination of thin section TEM combined with image averaging can identify the underlying structural defects associated with PCD	Computer-assisted TEM image averaging and 2-D alignment	Genetics	Sample from a single patient No quantitative validation Replication in other gene variants lacking Defined species-specific axonemal patterns and improved the resolution of the inner and outer dynein arms: Identified differences between human and C.  reinhardtii axonemes (e.g. ODA structure, number of RS) Images confirm that most of the inner arm dyneins and a portion of the dynein regulatory complex are not assembled in the PCD patient Supported PCD TEM interpretation	Sample from a single patient No quantitative validation Replication in other gene variants lacking
Jackson, 2022 [32]	Confirmed PCD patients (n=57), non-PCD patients (n=521)	IF	Determine if automated IF image analysis discriminates true “present” (non-PCD) from true “absent” (PCD) cilia proteins	Automated IF image analysis	Visual inspection of IF images by expert	Automated IF image analysis discriminated between PCD and non-PCD: Percentage overlap of target staining to control staining was significantly lower in PCD compared to non-PCD images (p<0.0001): o PCD group: median overlap of 0.8% o Non-PCD group: median overlap of 76.9% With a cut-off of 35% overlap: o Sensitivity: 95% o Specificity: 84%	Possibility of diminished discriminatory ability in case of intracellular tubulin stain
Liu Z, 2020 [31]#	Confirmed PCD patients (n=31) and healthy controls (n=6)	IF	Develop a super-resolution imaging-based toolbox to aid the diagnosis of PCD by: Developing a localisation map of PCD proteins by three-dimensional structured illumination microscopy Applying quantitative image analysis and a machine learning classifier to detect protein mislocalisation	The developed toolbox included: Developing a localisation map of PCD proteins by three-dimensional structured illumination microscopy (using wide field, deconvolution, and IF 3D-SIM) Applying quantitative image analysis and a machine learning classifier to detect protein mislocalisation Developing a high-throughput method for detecting motile cilia incoordination by rotational polarity	Genetics and IF visual inspection (21 antibodies)	Main findings included: Machine learning algorithm reliably distinguished cells from patients with PCD (null DNAH5 variants and partially localised DNAH5) from cells of healthy individuals o as well as from cells of other PCD patients (null CCDC39 and DNAH11 variants) Detection of 96-nm periodicity of GAS8 on motile cilia axoneme Rotational polarity analysis distinguished cells from patients with PCD, cystic fibrosis and healthy controls	Diagnostic comparisons only carried out for DNAH5 IF 3D-SIM could be further improved by implementing high-throughput staining and data collection methods such as multichannel microfluidic chips and automatic cell recognition/data analysis algorithms Rotational polarity can be used as a quantitative alternative in cases that cannot be diagnosed by other methods because assay time is longer due to requirement for ALI culturing that ensures the proper cell orientation for imaging For all methods discussed, a larger pool of PCD cases and healthy controls with different genetic background needs to be assayed before translation into clinical diagnostic pipeline

#: rows indicating studies describing AI-driven tools. HSVM: high speed video microscopy; TEM: transmission electron microscopy; CBE: ciliary beating efficiency index; CHD: congential heart disease; CBP: ciliary beat pattern; SCD: secondary ciliary dyskinesia; ALI: air–liquid interface; HNE: human bronchial nasal cells; nNO: nasal nitric oxide; HBE: human bronchial epithelial cells; AUC: area under the curve; ROC: receiver operating characteristic; ODA: outer dynein arm(s); MTD: microtubular disorganisation; CC: central complex; RS: radial spokes; IF 3D-SIM: three-dimensional structured illumination microscopy; IF: immunofluorescence.

17 studies focused on improving PCD diagnostic accuracy, through image analysis (n=6 were using AI-driven tools): eight records focused on HSVM of live ciliary function, seven records on TEM prediction of ciliary ultrastructural defects, and two records on immunofluorescence images to quantify presence or absence of structural ciliary proteins (Theme 1, table 1). No study identified new diagnostic testing strategies beyond those already recommended [8].

11 studies focused on improved PCD screening and disease characterisation; n=9 were using AI-driven tools (Theme 2, table 2). Three records showed data screening applications for PCD-related symptoms from population health records, insurance claims and structured questionnaires. Five studies used AI-driven tools: four to characterise computed tomography (CT) parameters from lung or craniofacial images to characterise clinical severity in PCD patients and one to assess the acoustic function of an electronic stethoscope not specific to PCD diagnostics but including two PCD patients in their study (meeting inclusion eligibility for our review).

TABLE 2.

Theme 2: characteristics of studies focusing on digital-automated or AI-driven tools used to better characterise primary ciliary dyskinesia (PCD) disease, screen clinical data or stratify PCD patients by genotype and phenotype

First author, year [ref.]	Study population	Application	Research question/aim of study	Digital/automated or AI method employed	Main findings	Limitations/challenges
Data mining applications for screening PCD-related symptoms
Burns, 2025 [37]#	Paediatric health insurance records (USA, n=1.38 million)	Disease characterisation (screening PCD-related symptoms/and claims)	Assess feasibility of a machine learning tool to screen a large administrative dataset to identify possible PCD cases	Random forest machine learning model trained on >40 clinical and procedural variables/claim codes (features) from a well-defined dataset of confirmed and highly likely paediatric PCD patients and healthy controls	Predicted 1 in 179 children likely to have PCD: Sensitivity 80–90% Top features: situs inversus, BAL, nasal biopsy	Validated only in insured US population; generalisability might be limited Potential data-linkage errors Dependent on insurance claims coding accuracy
Pfaff, 2020 [36]#	Confirmed PCD patients (n=22) and disease controls (n=203)	Disease characterisation (screening PCD-related symptoms)	Develop a tool to identify suspected PCD patients by searching text from the electronic health records (clinical notes)	Natural-language processing with Clinical Annotation Research Kit (CLARK) Did not utilise structured EHR data (i.e. ICD codes) that do not capture all data from clinical notes	CLARK can be used on clinical data of PCD to enhance the ability to define computable phenotypes that do not rely on structured EHR data, since there are no specific ICD diagnosis codes for PCD: Sensitivity 88% Specificity 100%	Dependent on text quality Small sample of PCD patients Classification largely relies on situs inversus; normal situs cases may be missed
Rother, 2015 [35]#	170 questionnaire responders: parents of children with confirmed PCD (n=24), CF (n=33), AS (N=27), PM (n=21), AB (n=23), PBB (n=18) and children without cough (n=24)	Disease characterisation (screening PCD-related symptoms)	Develop and test a questionnaire-based and data mining-supported tool to screen for selected pulmonary diseases based on respiratory symptoms	Computer-based data-mining program integrating eight machine learning classifiers Ensemble classifier combining outputs from all models to improve diagnostic prediction accuracy System trained to categorise new questionnaire responses and automatically suggest a likely diagnosis	For all diagnoses: 168 out of 170 patients received a correct diagnosis by the computer system Sensitivity:98.8% For PCD: 24 patients received a correct diagnosis Sensitivity: 100%	Small sample size Unclear confirmation of PCD diagnosis Single-centre study Participants contacted randomly during the hospital stay or during visits in the outpatient department (100% questionnaire return rate) while PCD participants mostly contacted via phone through patient support group (66% return rate)
Application in medical imaging to enhance clinical characterisation
Molica, 2024 [42]	Confirmed PCD patients (n=141)	Disease characterisation (medical imaging)	Assess an automated method to assess bronchus and airway (BA) dimensions in PCD	LungQ-BA Automated CT morphometry software to analyse bronchus–artery (BA) pairs	BA-LungQ appears to be an efficient and sensitive tool to describe airway characteristics and detect abnormally widened bronchi in children and young adults with PCD Detected widened bronchi in 57.9% of 9256 assessed BA pairs	Only 77 CTs (from 44 children) out of 141 children met the LungQ requirement of maximum slice thickness 1.5 mm Lack of direct comparison with manual evaluation Need for standardised protocols for wider adoption
Duan, 2023 [40]#	Confirmed PCD patients (n=32) and healthy controls (n=32)	Disease characterisation (medical imaging)	Utilise cranial CT scans to distinguish PCD from non-PCD based on differences in the sinus, middle ear mastoid, and fourth ventricle	Deep-learning CNN radiomics model trained on sinus and middle-ear features	Deep learning-based cranial CT imaging models can accurately screen for PCD Swin-Transformer, ConvNeXt and GoogLeNet training models displayed an AUC ≈0.94 Heat map revealed the differences in the sinus, middle ear mastoid and fourth ventricle between PCD and control group	Small dataset Single centre study Lack of validation dataset Large number of cranial CT pictures may be required
Duan, 2022 [38]#	Confirmed PCD patients (n=31) and controls (n=30) with CT of head and sinuses	Disease characterisation (medical imaging)	Analyse cranial CT characterisation using deep learning to differentiate otitis media (OM) caused by OM with effusion (OME) and PCD	Deep-learning CNN radiomics model trained on sinus and middle-ear features	DL-based cranial CT radiomics can accurately distinguish OM caused by OME from that caused by PCD Accuracy 0.887–0.904 depending on model depth	Small dataset Single-centre study CT layer thickness of temporal bone was 0.65 mm. Images of another layer thickness were not reported or studied
Hoang-Thi, 2018 [41]	Confirmed PCD patients (n=62)	Disease characterisation (medical imaging)	Develop an automated CT score to provide quantitative assessment of lung structure and assess correlation with lung function (FEV1 and FVC) in PCD	Automated CT image scoring method based on histogram characteristic analysis and threshold-based quantification of high-attenuating lung structures (attenuation value compared to threshold on CT)	Findings suggest automated density-based CT scoring, together with histogram characteristic analysis, is feasible in PCD and correlates well with lung function: CT scores based on fixed thresholds showed weaker negative correlations than did CT-density scores based on adapted thresholds taking into account histogram characteristics CT-density score based on MLD+1 sd had the best negative correlation with FEV1 (R= −0.68; p<0.001) FVC (R= −0.71; p<0.001) Automated score value was significantly different in the lower and upper portions of the lung, with higher score values in the lower part, consistent with the reported lower lung predominance of bronchial abnormalities in PCD	Retrospective study Unstandardised CT methods Technical limitation: pulmonary vessels are incorporated into the high-attenuating lung volume, which does not therefore correspond only to the diseased lung 11 participants not confirmed as PCD using diagnostic testing (had only clinical diagnosis as Kartagener's)
Menard, 2022 [43]#	Paediatric lung disease cohort (n=38) Underlying diagnoses were asthma (n=13), cystic fibrosis (n=8), acute viral bronchiolitis (n=9), tuberculosis (n=2), PCD (n=2), alveolar haemorrhage (n=1), croup (n=1), pneumonia (n=1) and chronic chest congestion (n=1)	Disease characterisation (auscultatory findings)	Assess AI-based electronic stethoscope for lung sound pattern recognition versus standard stethoscope auscultation	Wireless stethoscope and dedicated application employing machine-learning acoustic signal classifier based on StethoMe algorithms Concordance with standard auscultation by paediatric pulmonologists was assessed	The feasibility of the electronic stethoscope was demonstrated: Compared to standard auscultation, the electronic stethoscope was considered:   o worse in 11%   o identical in 55%   o better in 29% and   o much better in 5% Standard auscultation and manual electronic auscultation were concordant auscultatory findings in 27 out of 38 (71%) cases	Extremely small PCD sample Unclear confirmation of PCD diagnosis Larger replication studies are required
Pifferi, 2025 [46]#	Confirmed PCD patients (n=56) and healthy controls (n=50)	Genotype–phenotype correlation	Develop a predictive tool (decision tree) to identify HYDIN-specific ciliary motion signatures via HSVM and select patients that should be evaluated in detail for HYDIN variants	Principal component analysis of ciliary motion parameters (hierarchical clustering) and neural network model	The decision tree algorithm to identify patients with HYDIN genetic variants was based on the presence of stiff and circular motion with normal CBF >AUC: 98% for HYDIN classification	Single-centre study with small training set Small number of PCD patients with HYDIN genetic variants (n=8) No external validation High-quality HSVM required
Pifferi, 2024 [45]#	Confirmed PCD patients (n=212)	Genotype–phenotype correlation	Use by soft computing analysis to predict PCD genotype groups based on ciliary TEM and HSVM parameters	Principal component analysis (hierarchical clustering) and neural network model	Identified distinct motion clusters by gene groups PCA can predict pheno-geno clusters Model accuracy ≈98% for genotype group prediction	Single-centre study with small training set Unclear selection of gene groups No external validation
Walton, 2023 [44]#	Confirmed PCD patients (n=4) and healthy control (n=1). Additional >36 000 micrographs of C.  reinhardtii axonemes	Disease characterisation (axonemal structures)	Employ cryo-electron microscopy and artificial intelligence-enabled structure prediction to predict the structure of the 96-nm modular repeats of axonemes in the flagella of C.  reinhardtii and human respiratory cilia	Image analysis AI-based prediction (using AlphaFold2, SWISS – MODEL) Cryo-EM maps (Chimera X or COOT) for building models of axonemal protein structures	The employed methodology can: Map axonemal protein interactions Assess the effect and pathogenicity of disease variants by linking structural loss to variant type Identify new candidate genes in unsolved PCD cases Assess how the loss of individual docking factors can selectively eradicate periodically repeating structures	Highly specialised tool which limits wide adoption Replication in other gene variants lacking

#: rows indicating studies describing AI-driven tools. AI: artificial intelligence; BAL: bronchoalveolar lavage; EHR: electronic health record; CF: cystic fibrosis; AS: asthma; PM: pneumonia; AB: acute bronchitis; PBB: protracted bacterial bronchitis; CT: computed tomography; CNN: convolutional neural network; CBF: ciliary beat frequency; AUC: area under the curve; FEV1: forced expiratory volume in 1 s; FVC: forced vital capacity; MLD: mean lung density; TEM: transmission electron microscopy; HSVM: high-speed video microscopy; PCA: principal component analysis.

A graphical summary of promising applications of AI and digital-automation tools in PCD is presented in figure 2.

FIGURE 2.

Summary of the promising applications of digital-automated and artificial intelligence (AI)-driven tools in primary ciliary dyskinesia (PCD). TEM: transmission electron microscopy.

Risk of bias assessment

Using the modified CLARITY tools for cohort and case–control studies, most included studies demonstrated low overall risk of bias. Across domains, case selection, exposure and outcome assessment were generally robust. However, quantitative analyses were often limited, preventing adjustment for potential confounding variables. The predominantly single-centre, exploratory designs and small sample sizes restricted external validity rather than introducing systematic bias. Overall, methodological quality was acceptable, with low risk of bias across both diagnostic and disease characterisation studies (supplementary table S2).

Qualitative synthesis Theme 1: automated digital tools or AI-driven tools that are used to enhance ciliary function, ultrastructure or immunofluorescence analysis for PCD (table 1)

Application in ciliary function analysis

Tools can automate CBF and CBP measurements following HSVM of live ciliary function. Eight studies showed novel indices of waveform and beat pattern characteristics, reporting high diagnostic accuracy reduced analysis-time and observer variability compared to manual assessment.

Papon et al. [15] developed custom software for quantitative HSVM analysis. The software distinguished PCD from healthy control samples with high sensitivity and specificity by analysing objective motion metrics such as distance travelled by the cilium tip weighted by the percentage of beating ciliated edges, while also markedly reducing interobserver variability. Although not assessed in PCD patients, Sears et al. [16] studied ciliary beating waveform (using nasal and bronchial samples from healthy individuals) and showed consistent and repeatable measurements across different normally functioning samples. Among other CBP parameters, the model quantified shape of the waveform, curvature (bending) along the cilium, shear angle and velocity that would otherwise be subjectively assessed by a human observer. Quinn et al. [17] similarly tested an algorithm to computationally discriminate normal versus abnormal (dyskinetic) ciliary function in HSVM data. Whilst test accuracy was high and analysis-time was reduced, cross-validation using other sourced video data was not evidenced. Quinn et al. used expertise to select appropriate quality videos for analysis which would improve algorithm accuracy and therefore could cause underperformance of this tool in other settings (particularly with limited resources and training/expertise). Digital tools will reduce subjectivity in providing qualitative beat pattern analysis measures for HSVM testing that will discriminate PCD from secondary dyskinesia, but open-access tools and validatory work on different sample types is currently not evidenced to facilitate this. Bottier et al. [18] quantitated microbead tracking, in patients’ in vitro cultured airway samples, to derive a normative motion index termed ciliary beating efficiency (CBE) as a functional measure of mucociliary fluid flow. Whilst not validated in PCD, this CBE method was independently used to characterise disrupted mucociliary fluid flow in samples from patients with chronic rhinosinusitis [19]. Open-source platforms such as Cilialyzer [20] allow the user to not only measure CBF but also carry out particle tracking analysis within user-defined regions, but validation studies are needed to demonstrate usability and accuracy for PCD diagnostics.

Pifferi et al. [21] demonstrated a neural network-based machine learning model that distinguished rotation, migration and mucociliary clearance differences between cultured airway spheroids designated PCD and non-PCD samples based on abnormal or normal TEM; a normal TEM does not exclude PCD. Chioccioli et al. [22] used an unsupervised k-means clustering machine learning algorithm to differentiate PCD with disease-causing variants in HYDIN and DNAH11 based on reduced beat amplitude and curvature and healthy controls; cases that lack of discernible TEM defects and without other function testing a PCD diagnosis could be missed. These algorithms are not publicly available by open access, and further validation and expanded testing strategies are needed.

Application in ciliary ultrastructure analysis using TEM

PCD can be diagnosed based on expert TEM analysis of ciliary ultrastructure but is subjective and requires a sufficient number of axonemes for quantification [6]. Seven studies used automated digital and/or AI-driven tools to assess TEM images with high sensitivity and specificity. Early work using computer-assisted alignment and image averaging algorithms by O'Toole et al. [23] and Funkhouser et al. [24] aiming to standardise dynein arm assessment was found to outperform manual evaluation and reduce subjectivity in TEM analysis [25]. Semi-automated pipelines such as the PCD-Detect, developed by Shoemark et al. [26], applied alignment, averaging and unsupervised clustering techniques to classify ultrastructural features, which detected subtle central complex defects (caused by pathogenic HYDIN variants) better than human analysis. PCD-Detect to analyse TEM was independently validated in patients with bronchiectasis with inconclusive PCD genetic outcome [27]. Pinto et al. [28] and Bottier et al. [29] subsequently developed PCD-AID, an AI-driven tool trained to detect cilia images of appropriate quality and classify hallmark defects diagnostic of PCD with high accuracy and analysis time-saving. An AI-driven deep-learning PCD-Quant tool for quantifying ultrastructural components reduced examination time five-fold whilst remaining diagnostically sensitive [30].

Most systems described require open-accessibility and external validation to facilitate widespread diagnostic implementation.

Application in ciliary ultrastructure analysis using immunofluorescence imaging

Absence or mislocalisation of key ciliary proteins (as assessed by immunofluorescence imaging) can support a PCD highly-likely diagnosis, particularly in cases lacking diagnostic confirmation by TEM or genetics but are characterised by a combination of indicative clinical symptoms, abnormal ciliary function and abnormal nNO. Two studies implemented automated digital and/or AI-driven methods to quantitatively analyse immunofluorescence images.

Liu et al. [31] created a super-resolution imaging toolbox of 21 commercially available PCD-relevant antibodies combined with a machine learning classifier. This approach enabled high-resolution quantification of protein distribution patterns that may be difficult to discern by standard microscopy. However, the requirement for specialised equipment and expert operators limits scalability in routine or low-resource settings. A higher-throughput alternative developed by Jackson et al. [32] achieved high testing sensitivity and specificity when using an automated ImageJ-based analysis platform, to quantify the percentage presence or absence of immunofluorescence-labelled ciliary proteins in PCD cases versus non-PCD controls, based on automated identification of “mask” analysis regions, identified by good-quality tubulin-immunofluorescence-labelling of cilia. Whilst quantitative immunofluorescence analysis is preferable to subjective analysis, cross-validation of either of these tools remains untested.

Qualitative synthesis Theme 2: automated digital tools or AI-driven tools used to better characterise PCD disease, screen clinical data or stratify PCD patients by genotype and phenotype (table 2)

Data mining applications for screening PCD-related symptoms

The median age of PCD diagnosis in Europe has presently been reported to be 5.3 years [33], and similar estimates have been reported in North America [34]. Three studies used machine learning approaches to data mine for patients with indicative symptom patterns in questionnaire data or electronic health records (EHRs) and insurance records.

Data from a PCD-relevant symptom questionnaire, developed by Rother et al. [35], was tested with a “classifier” algorithm and distinguished PCD from other respiratory conditions with high sensitivity, although evidence of diagnostic certainty was not shown.

Pfaff et al. [36] developed an open-source phenotyping natural language model software called the “Clinical Annotation Research Kit” (CLARK), to stratify patients for several diseases, by extraction of clinical features from text sources. In the absence of a specific ICD-10 diagnosis code for PCD, the usefulness of this approach to screen EHR is promising; however, PCD will likely be missed irrespective of diagnostic pathway (or tool use), and diagnosis of situs inversus (seen in ∼50% of PCD cases) is a data bias that could unwantedly overlook PCD cases with situs solitus (normal). A preprint by Burns et al. [37] described an AI-driven random forest plot model, trained on data from highly-likely/proven PCD versus “healthy” paediatric patients. When applied to a screening cohort of 1.38 million US health insurance claims (<18 years), 1:7705 were predicted to have PCD, which is similar to estimates by Hannah et al. [5]. The observed sensitivity was high using the statistically most predictive features: situs inversus, bronchoalveolar lavage, nasal biopsy, inhaled hypertonic saline, airway clearance, antibiotics, pulmonary diagnoses, endoscopic airway exam, airway cough and chronic sinusitis. Although promising, this screening tool was only validated in insured US children, and limited by the presumed accuracy of drug prescription, insurance codes and automated data linkage.

Application in medical imaging to enhance clinical characterisation

Limited single-centre studies have explored AI-driven tools to improve quantitative analysis of CT-based imaging to characterise PCD disease severity.

Duan et al. [38] applied a deep learning neural network radiomics [39] model to differentiate radiological features of otitis media with effusion in PCD patients compared to non-PCD patients. PCD screening was attempted on cranial CT scans, using deep learning training models to identify differences in the sinus, middle ear mastoid and fourth ventricle scans of PCD patients versus non-PCD control group [40]. For both studies, sample sizes were small and results were unverified using independent centre images. Hoang-Thi et al. [41] carried out quantitative assessment of lung structural abnormalities in adult PCD patients using an automated CT analysis score to indicate disease severity. ∼18% of the patients lacked confirmatory PCD diagnostic testing, and pulmonary structures that do not correspond to PCD disease severity undermine this study. Automated digital quantitative analysis of bronchial and arterial dimensions in chest CT scans of PCD patients with bronchiectasis detected widened bronchi, although validation of this tool compared to subjective visual scoring and qualitative visual assessment was not reported [42].

Automated digital and AI-driven imaging tools offer objectivity and quantitative feature extraction, which allow for reproducibility. Larger multicentre studies and external validation are required before clinical integration as current evidence is limited. A study capturing use of an AI-driven electronic stethoscope met our inclusion criteria due to the inclusion of two PCD patients. There is insufficient evidence for its specific utility in PCD [43].

Application in characterising axonemal abnormalities and informing genotype and phenotype correlations

AI-driven tools have been applied to PCD ciliary ultrastructure abnormalities and in some studies compared by specific PCD genes. Walton et al. [44] applied cryo-electron tomography with AI-driven structural prediction to map 96-nm axonemal repeats and successfully predicted loss of ultrastructural elements in PCD patient samples with biallelic pathogenic CCDC39 and CCDC40 variants, although it is not an open resource tool or cross-validated. Pifferi et al. [45] used principal component analysis followed by a neural network algorithm to predict PCD ciliary function characteristics and genotype, when filtered by TEM ultrastructural defect. The model accurately correlated genotype with TEM; however, it applied prior published knowledge introducing assumption bias and would likely miss atypical PCD cases in other settings. In a follow-up study, a decision tree model was developed to select patients for HYDIN-specific genetic testing, based on abnormal stiff and rotational CBP combined with normal CBF (by HSVM) [46].

Future directions: ERS CONNECT CRC focus group narrative

Focus group question 1: what are the practical steps and ethical considerations needed for integrative use of automatic digital and AI-driven tools; to improve patient screening and PCD diagnostic accuracy, and to better stratify diagnosed patients for tailored care and research?

Integrating automatic digital and AI-driven tools into PCD screening, diagnosis, and patient stratification offers significant opportunities but requires planning and ethical oversight. Current automatic digital and AI-driven tools in PCD are largely limited to small, single-centre studies, which restricts their use due to lack of external validation. Bias and representativeness of data are major risks, e.g. where records held do not reflect the whole population being studied. Models and tools trained on limited or skewed datasets potentially systematically underperform by excluding atypical patient clinical characteristics, genotypes and phenotypes. Pre-imposed assumptions should be avoided when screening healthcare records, particularly due to a lack of universal disease-specific identifiers for PCD. Evidence-based and well-defined diagnostic test outcomes (e.g. the BEAT-PCD TEM consensus reporting guidelines within the 2025 ERS-ATS PCD diagnostic guidelines) are separately important for supervised model training. Integration planning must define clear workflows for how AI outputs should be interpreted, such as disease risk scores or thresholds or automated tool classifiers with clinical decision-making governed by the clinician overseeing patient care. Another type of bias could be automation bias (i.e. over-relying on automated systems), which could make users less vigilant of potential errors and requires consensus safeguarding strategies.

In preparation, a staged validation pipeline is essential, beginning with local proof-of-concept studies and progressing to external tool/model validation using data from diverse populations and healthcare systems. To improve patient screening and advance multicentre collaborations, pooling of single-centre datasets to create larger, more globally representative datasets is critical. Open-source algorithms and shared data repositories could be prioritised to enable and facilitate cross-validation, re-training and further development of better algorithms with less data bias. Alternatively, federated machine learning approaches could be used, where a central server can train a shared AI model across multiple other servers or devices without disclosing raw sensitive data and removing barriers relating to data privacy, e.g. such as those used during the COVID-19 pandemic, where AI tools analysed chest scans across international hospitals [47]. Food and Drug Administration (FDA) regulatory approval of validated AI-driven software as a medical device (SaMDs) is considered best practice [48].

Transparency and accountability of tool/model development must be ensured, clarifying responsibility for errors between developers and clinicians to safeguard against misuse. It is essential to also perform continuous model auditing for performance disparities across subgroups. Black-box or proprietary models can accurately identify complex patterns in patient data; however, lack of process transparency raises ethical concerns around model validation and performance auditing particularly in the context of rare diseases where data are limited by rarity (reviewed in He et al. [49]). The EU AI Act (2026) is the first global comprehensive legal framework to be used by national authorities and an EU AI Office that could enforce high fines for rule violations, focusing on data safety, fundamental human rights and transparency [50]. Use of biometric categorisation of protected characteristics or cognitive manipulation are prohibited. The EU AI Act considers AI systems that impact healthcare, education and law enforcement as high risk that must meet data governance standards and compliance rules with risk assessments and management strategies disclosed. The use of general purpose, e.g. large language, AI models also requires obtainment of copyright agreements and must comply with data transparency codes of conduct. Use of retrospective clinical imaging, video microscopy or genomic data for AI training may raise consent challenges to be considered. Despite the EU framework, further development and success of implementation will likely lag the rapid advancement of AI technologies and is also affected by geo-political issues.

Environmental impact is also a concern as large-scale AI training consumes significant energy, contributing to carbon emissions. However, there have been examples of models optimised for resource limited settings, especially those utilising smartphone-based systems [51, 52], as well as examples of deep neural networks able to overcome the limited data constraints that may characterise rare and paediatric diseases [53].

Focus group question 2: what methods/examples/considerations will support resource-limited areas to include use of automatic digital and AI-driven tools despite patient clinical/diagnostic data gaps?

Researchers in resource-limited settings may rely more frequently on poorly annotated medical datasets, resulting from both lack of expertise and financial resources. Previous studies have highlighted the potential of deep convolutional neural networks to correctly classify medical images despite limited labelled data [54]. Furthermore, where performance gains are marginal, simpler models should be favoured, and local systems rather than large virtual cloud-based models should be used to behave responsibly with regard towards global climate change. Differences in healthcare systems or diagnostic services (due to limited resources) likely mean that models validated in one context may not transfer seamlessly to others, due to missing types of data, uncertain or inaccurate data, necessitating adaptation and re-validation of models to bespoke use. This requirement is especially relevant for machine learning models that leverage EHR data for clinical or screening algorithms. Although previous studies have shown that patterns discovered in EHR data using machine learning are applicable to real-life settings, these models tend to capture the local EHR structure, and their generalisability outside the EHR(s) from which they were developed is limited [55]. Finally, prioritising agnostic and more flexible model designs over highly model- or data-specific approaches may improve generalisability and ease of implementation across different healthcare environments, including those with limited resources.

ERS CONNECT CRC focus group insights for both questions are summarised in table 3.

TABLE 3.

Summary of focus group insights

Focus group question	Reported strategies – actions
Focus group question 1: What practical steps and ethical considerations are needed for integrative use of artificial intelligence (AI) and automated/digital applications; to improve patient screening and primary ciliary dyskinesia diagnostic accuracy, to better stratify diagnosed patients for tailored care and research?	Establish multicentre collaborations Pool datasets across centres Implement a staged validation pipeline Integration planning Prioritise open-source tools and algorithms Avoid duplication of tools Regulatory compliance Ensure validation across diverse populations Maintain clinician oversight Introduce safeguards against misuse
Focus group question 2: What methods/examples/considerations will support resource-limited areas to include use of AI and automated/digital applications despite patient clinical/diagnostic data gaps?	Implement collaborative networks Leverage lightweight and energy-efficient model designs Rely on methods that perform well with limited or poorly-annotated data Adapt and re-validate models locally Prioritise agnostic model and data architectures

Discussion

Main findings

Digital automation and AI in PCD address three main objectives: identifying cases, diagnosis and research/discovery. Screening approaches, largely AI-driven and applied to electronic health records or administrative datasets, aim to identify individuals who may warrant further diagnostic assessment, but performance is context dependent and requires local validation. Diagnostic support tools are the most developed, with automated digital and AI-driven approaches applied to HSVM, TEM and immunofluorescence to standardise interpretation and reduce interobserver variability. Several studies reported reductions in analysis time compared with manual assessment, suggesting potential advantages within specialist services. Research-focused tools, including genotype–phenotype clustering and structural modelling, provide insights into disease mechanisms and variant interpretation but remain mostly research focused. All these potential uses of digital technology augment specialist workflows rather than replace expert judgement.

Strengths and weaknesses

This review synthesises automated digital and AI-driven approaches across the PCD diagnostic and research pathway. The literature shows that objective, quantitative methods can reduce variability in interpretation and may reduce specialist workload for selected tasks. A small number of tools are open-source or sufficiently documented to permit independent testing and use by other centres. However, most studies were single-centre, retrospective and small, with limited external validation and heterogeneous reporting of performance. Many models were developed and evaluated in case–control settings rather than real-world referral populations, limiting generalisability. Several AI-driven approaches relied on imperfect reference standards, with potential to propagate diagnostic uncertainty. There is currently no perfect PCD reference test; ∼30% of PCD lack discernible TEM defects [6], and ∼25% lack a genetic diagnosis [56], not accounting for the likely missed variants that are non-coding, copy number, in-frame deletions and aberrant splicing [57]. Without a definitive reference, misclassification of training data for neural network-based models will be likely. Restricted access to code and training data limits reproducibility, and few studies evaluated clinical impact, workflow integration or cost-effectiveness.

Clinical and research context

Within current diagnostic pathways for PCD, the most likely role for digital and AI-driven tools is in supporting specialist diagnostics by standardising interpretation and reducing time spent on HSVM, TEM and immunofluorescence analysis. Tools that assist with image quality control and triage of diagnostically usable data may be more straightforward to integrate into existing workflows. Screening approaches based on electronic health records or administrative data may assist case-finding but needs careful local adaptation because healthcare systems and the completeness of data are widely varied. Research tools, including genotype–phenotype clustering and structural modelling, are currently more relevant to discovery and variant interpretation than routine clinical decision-making. Adoption beyond expert centres will depend on usability and ability to combine these with existing infrastructure and pathways.

Future directions

Future work should prioritise multicentre validation across diverse populations and healthcare settings, with prospective evaluation integrated into routine diagnostic pathways where feasible. Standardised reporting of performance and greater transparency around code and training data would improve reproducibility and comparability. Development should focus on tools that integrate with existing workflows and perform reliably with heterogeneous datasets to support implementation beyond the centres in which they were developed and expert centres. Alongside discriminative models reviewed here, generative AI approaches (e.g. large language models) may have future roles in research support (e.g. data annotation, literature synthesis and documentation), but there is currently insufficient evidence to support clinical use in PCD diagnosis or management. Integration of digital tools and AI into clinical care for PCD should proceed with rigorous validation and appropriate governance and regulatory oversight. Ethical and data governance considerations, including anonymisation, consideration of accuracy and bias, should be addressed alongside technical development.

Supplementary material

Please note: supplementary material is not edited by the Editorial Office, and is uploaded as it has been supplied by the author.

Supplementary material

01694-2025.SUPPLEMENT