New CT-based dural ectasia criteria using machine learning to diagnose Marfan and Loeys-Dietz syndromes

Abstract

Aims

Dural ectasia (DE) is a criterion for Marfan syndrome (MFS) diagnosis. However, there is no agreement on its diagnosis. We identified new DE criteria on CT-scan imaging using machine-learning (ML) and aimed to evaluate their performance for MFS and Loeys-Dietz detection.

Methods and results

MFS patients with FBN1 pathogenic variant, who underwent a complete CT-scan were included and matched 1:1 with controls. The conventional criteria for DE were assessed, as well as new criteria identified by ML (generalized linear model and random forest), regarding MFS diagnostic performance. The new DE criteria were then evaluated for identification of Loeys-Dietz patients with TGFßR1/2 and SMAD3 pathogenic variants. A user-friendly online interface was finally developed to predict the probability of MFS or Loeys-Dietz diagnosis.

Between November 2010 and January 2017, 93 MFS patients and their matched controls were included. Mean age was 39 ± 13 years with 45% of women. Conventional definitions of DE showed poor diagnostic performance for MFS: AUC 0.68 (0.61–0.74).

The new DE criteria, based on simple CT-scan measurements (antero-posterior diameter of the spinal canal and scalloping) on vertebrae L1, L2 and S1, achieved AUC 0.84 (0.69–0.94). These criteria outperformed the Ghent-criteria for MFS diagnosis. Moreover, in 46 patients with TGFßR1/2 and 40 with SMAD3 pathogenic variants, the new criteria also achieved good diagnostic performance (AUC 0.83 (0.73–0.90) and 0.80 (0.70–0.88) respectively).

Conclusion

New CT-scan DE criteria showed good performance for detecting MFS or Loeys-Dietz syndromes. Given the importance of early diagnosis, these new simple criteria offer a promising screening tool.

Graphical Abstract

Graphical Abstract.

A new dural ectasia (DE) algorithm was developed from CT-scan imaging using a machine-learning approach based on four measurements of three vertebrae. These new DE criteria outperformed previously published DE criteria and showed excellent diagnostic performance for MFS and LDS patients. They might contribute to earlier referral of patients for proper management.

Introduction

Marfan Syndrome (MFS) is a genetic heritable thoracic aortic disease (HTAD), leading to cardiovascular, musculoskeletal, ophthalmic, pulmonary and neurological manifestations, one of which is dural ectasia (DE).¹ The altered fibrillin-1 (FBN1) protein in MFS may weaken the dura mater, leading to its expansion.

MFS diagnosis is based on genetic results, clinical signs, and family history, according to the Ghent criteria, which were revised in 2010 (Ghent-2 criteria).² Pathogenic FBN1 variants are identified in approximately 85–95% of patients fulfilling the revised Ghent criteria.^3,4 Once diagnosed, patients undergo regular cardiovascular imaging, to monitor aortic dilation.^5,6 Early diagnosis is crucial in order for patients to benefit from appropriate treatments and follow-up, since thresholds for surgical referral are lower in case of confirmed MFS than in the general population.^5,6 The main aim is to avoid aortic dissection.^7,8 Among the 13 systemic Ghent-2 criteria, several may be user-dependent, particularly skeletal signs. Additionally, while FBN1 pathogenic variants play a crucial role for diagnosis, proper genetic assessment requires an additional period of time, and some MFS patients do not have an identified mutation.⁹ This may be due to lack of sensitivity of the method used or to pathogenic variants in other genes than FBN1. Regarding Loeys-Dietz syndrome (LDS) genetic background, the more frequent identified pathogenic variants are on TGFBR2, TGFBR1, and SMAD3 genes (55–60%, 20–25% and 5–10%, respectively). Among patients presenting with Marfan-like features, pathogenic variants in TGFBR1/2 are identified in approximately 3–8% of cases, while SMAD3 variants are identified in 0.7% to 3.4% of cases.^10,11 Patients with LDS require additional genetic analysis.^12,13 Multigene panel that includes FBN1 and LDS genes may thus be of interest but are more costly.¹⁴ Hence, diagnosis may sometimes only be confirmed 2 to 4.5 years from the time of the first clinical manifestation.¹⁵

We aimed to determine if a new definition of dural ectasia (DE) may diagnose MFS. Published criteria for determining DE are variable, with wide heterogeneity in both imaging techniques and measurements approaches, and have demonstrated poor specificity,^16–19 leading to a reduced weight of DE in the revised Ghent criteria.² DE can be measured from a CT-scan of the chest, abdomen, and pelvis (CT CAP) allowing the visualisation of the lumbar spine. A reference CT-scan (or MRI) is recommended to adequately visualize the entire aorta when MFS is suspected, thus covering the lumbar spine, and DE assessment would require no additional examination.^5,6

Imaging multiple vertebrae generates a large number of parameters, which may complicate data interpretation. Herein, we used a machine learning (ML)-based approach to establish an algorithm for new DE criteria, based solely on three vertebras and four simple quantitative measurements (antero-posterior diameter of the spinal canal and scalloping) from the CT-scans already performed. Our primary objective was to define a new DE algorithm that combines high performance for MFS detection with a limited number of required measurements, thereby facilitating clinical application. Our secondary objective was to assess the performance of this new algorithm for LDS diagnosis.

Methods

Study population

Our study included 93 patients with identified FBN1 pathogenic variant, referred to as MFS patients, and 86 patients with pathogenic variants in the TGFß pathway (46 in the TGFβR1 or 2 genes and 40 in the SMAD3 gene, referred to as LDS patients) along with age- and sex-matched controls. All MFS and LDS patients of this study were included in the French reference center for MFS and related-syndromes at Bichat Hospital, Paris, France, as part of a comprehensive multidisciplinary Aortic Team evaluation, the organisation of such expertize having been previously published.²⁰ Patients with suspected MFS systematically underwent an initial reference CT scan to assess the aorta, as recommended.^5,6 Sequence variant reporting was conducted according to Human Genome Variant Society nomenclature in the reference genetic laboratory.²¹ This study was conducted in accordance with the principles of the Declaration of Helsinki and meets the STARD and TRIPOD requirements.²² The study obtained the Human Study IRB approvals (n°1229912 and n°2207326) and was sponsored by the French Society of Cardiology and registered in the Institut National des Données de Santé (INDS) (registration number 0810221118) as required by local regulations.

We retrospectively included consecutive patients according to the following criteria (Figure 1): adult patients diagnosed with MFS or LDS diagnosis with an identified pathogenic variant in the FBN1, TGFβR1 or TGFβR2, or SMAD 3 genes. Exclusion criteria were patients under 18 years of age, patients without an injected CT of the chest, abdomen, and pelvis (CT CAP) allowing analysis of DE, and patients with spinal surgery that hindered proper analysis of the lumbosacral cord.

Figure 1.

Diagram of inclusion. The three cohorts of MFS patients (with FBN1 pathogenic variant) and Loeys-Dietz syndrome patients (TGFβR1 or TGFβR2 and SMAD3 pathogenic variants) were matched 1:1 with age and sex controls. * These patients originate from the same cohort.

Age- and sex-matched controls were selected from the database of CT-scans conducted at the same center (Bichat Hospital), among adult patients with an injected CT CAP enabling the analysis of DE, and without any known aortic or spine diseases. Clinical data from MFS and LDS patients were gathered during the multidisciplinary visit and data from control patients were obtained from their medical records.

We initially aimed for a minimum of 60 patients in each group to achieve power of analysis of 80% with an anticipated ROC curve AUC of 0.85 and P value of 0.05 (Power calculations conducted with the pROC package). Therefore, the final population of 93 patients in each group ensures appropriate power for our analysis.

CT-scan DE criteria determination

Several DE measurements derived from the literature were performed on the reference CT CAP images, on the sagittal plane:

• the dural sac ratio (DSR), defined as the ratio between the antero-posterior diameter of the spinal canal and the antero-posterior diameter at the center of the vertebra. Both measurements were taken perpendicular to the dural sac axis and the vertebra, with zoom and alignment facilitated by a multiplanar reconstruction mode (MPR) (Figure 2A and B).

• vertebral scalloping, describing a configuration change in the vertebra resulting from central vertebral core erosion. The scalloping value was determined using the measurements of the superior (A) and inferior (C) vertebral plateau and the antero-posterior diameter at the center of the vertebra core (B), with the following formula: (A + B)/2-C. Each measurement was performed by zooming on each vertebra and aligning it in MPR mode at the center of the vertebra and perpendicular to its axis (Figure 2A and B). Measurements were performed by two readers blinded to the clinical data.

Figure 2.

CT scan imaging of dural ectasia parameters. (A): measurement of the upper (A) and lower (C) plateaus of the L4 vertebra, of the anteroposterior diameter at the center of the vertebra (B) and of the spinal canal (Red) in a MFS patient. Measurements were taken in sagittal section (panel 1), with prior alignment centred on the vertebra in transverse (panel 2) and frontal (panel 3) sections. (B): measurements of the upper (A) and lower (C) plateaus of the L4 vertebra, the anteroposterior diameter at the center of the vertebra (B) and the spinal canal in a control patient. Measurements were taken in sagittal section (panel 1), with prior alignment centred on the vertebra in transverse (panel 2) and frontal (panel 3) sections.

Statistical analysis & machine learning (ML)

Analyses used means (standard deviation), and t-test or Wilcoxon test depending on the distribution for the comparison of continuous variables, and numbers (percentage, %), and Chi-2 or Fisher exact test for the comparison of discrete variables.

This study was conducted according to the TRIPOD guidelines (see Supplementary material)²³ The R software with the tidyverse environment²⁴ was used to analyse the database formed by merging the clinical data and the DE CT-scan measurements. Missing data were minimal, accounting for only 0.8%, in the FBN1 pathogenic variant database. Missing values were imputed using missranger. The main objective to be predicted was positivity for MFS diagnosis as ascertained by a pathogenic variant in the FBN1 gene.

To build the DE algorithm, all the vertebra measurements which demonstrated a t-test P value below 0.2 between MFS and control patients were selected, to avoid the exclusion of covariates. A glmmulti-mediated approach using the glmulti package²⁵ selecting by the Akaike information criterion, was further used, extracting the most relevant variables. These were then used to build preliminary GLM and RandomForest models to extract their importance (arbitrary scale) in each model (see Supplementary data online, Table S1).

A standard training/testing set approach was adopted on the MFS cohort (80/20 split). The artificial intelligence classifiers were chosen based on the literature, considering both their performance and technology, each representing a different approach to extracting relevant information from the data.^25,26

The performance of generalized linear model (GLM), K-nearest neighbour (KNN), support vector machine (SVM), random forest (RF) and boosted trees (C5,0) was compared. Candidates with the best performance were selected based on several indicators: the area under the curve (AUC) of the Receiver Operating Characteristic (ROC) curve, highlighting overall performance, was evaluated according to established metrics²⁷; the sensitivity, proportion of variant carrier patients correctly identified within the population of variant carrier patients; the specificity, proportion of true negatives correctly identified within the controls; the P value compared with the no information rate [P-value (Acc > NIR)]; the J index and the Brier class.

The risk of overfitting was assessed by implementing a decoy strategy, where the outcome variable was scrambled, and the entire machine learning strategy was re-executed. Standard statistical tests (Student’s t-test and Spearman’s correlation) were also performed with R. The significance level was set at 0.05.

Results

Patients’ characteristics

Between November 2010 and January 2017, 93 MFS patients with proven FBN1 pathogenic variants were enrolled in this study and matched with 93 controls, as well as 86 LDS patients (46 with TGFßR1 or 2 pathogenic variants, and 40 with SMAD3 pathogenic variants), all matched with controls (Figure 1). The mean age of the FBN1 population was 39 ± 13 years old, with 55% of patients being male, without a significant difference between MFS and control patients, indicating appropriate matching (Table 1). All cardiovascular risk factors were more prevalent in controls than in MFS patients. Regarding CT-scan parameters of DE, they were all significantly increased in MFS patients as compared with controls, except for the anteroposterior diameter at the centre of the vertebra for L1, L2, L3 and L4.

Table 1.

Characteristics of MFS (FBN1 pathogenic variant) patients

Parameter	Marfan Patients n = 93	Control Patients n = 93	P value
Male Sex (%)	51 (55)	51 (55)	1
Age (year)	39.1 ± 13.4	39.2 ± 13.3	0.97
BMI (kg/m2)	23.4 ± 4.9	24.6 ± 6.6	0.18
Total Body Surface Area (m2)	2 ± 0.3	1.8 ± 0.2	<0.001
Systolic Blood Pressure (mmHg)	128.4 ± 14.3	124.3 ± 23.6	0.17
Diastolic Blood Pressure (mmHg)	73 ± 8.7	74.9 ± 13.6	0.28
Smoking (%)	4 (4.3)	23 (24.7)	0
Hypertension (%)	8 (8.6)	13 (14)	0.14
Dyslipidemia (%)	2 (2.2)	5 (5.4)	0.17
Diabetes (%)	0	11 (11.8)	<0.001
Coronary artery disease (%)	0	6 (6.5)	0.01
L1 Spinal Canal Antero-Posterior Diameter (mm)	19 ± 2.6	16.4 ± 2.1	<0.001
L2 Spinal Canal Antero-Posterior Diameter (mm)	18.3 ± 2.5	16 ± 2.1	<0.001
L3 Spinal Canal Antero-Posterior Diameter (mm)	17.1 ± 2.9	15.3 ± 2.4	<0.001
L4 Spinal Canal Antero-Posterior Diameter (mm)	18.6 ± 3.5	15.6 ± 2	<0.001
L5 Spinal Canal Antero-Posterior Diameter (mm)	20.9 ± 4.9	16.7 ± 3.7	<0.001
S1 Spinal Canal Antero-Posterior Diameter (mm)	19.6 ± 5.8	14 ± 3.1	<0.001
S2 Spinal Canal Antero-Posterior Diameter (mm)	14.7 ± 7.7	8.5 ± 2.2	<0.001
L1 Antero-posterior diameter at the center of the vertebra (mm)	29.1 ± 3.9	28.9 ± 2.9	0.64
L2 Antero-posterior diameter at the center of the vertebra (mm)	30.1 ± 3.8	30.1 ± 2.9	0.94
L3 Antero-posterior diameter at the center of the vertebra (mm)	32 ± 3.7	31.2 ± 2.8	0.12
L4 Antero-posterior diameter at the center of the vertebra (mm)	31.9 ± 3.6	31.8 ± 2.7	0.89
L5 Antero-posterior diameter at the center of the vertebra (mm)	30.6 ± 4.1	31.6 ± 3.1	0.05
S1 Antero-posterior diameter at the center of the vertebra (mm)	20.7 ± 5.1	25.1 ± 2.6	<0.001
S2 Antero-posterior diameter at the center of the vertebra (mm)	10.8 ± 5.5	15.9 ± 2.9	<0.001
L1 Dural Sac Ratio	0.7 ± 0.1	0.6 ± 0.1	<0.001
L2 Dural Sac Ratio	0.6 ± 0.1	0.5 ± 0.1	<0.001
L3 Dural Sac Ratio	0.5 ± 0.1	0.5 ± 0.1	0.001
L4 Dural Sac Ratio	0.6 ± 0.1	0.5 ± 0.1	<0.001
L5 Dural Sac Ratio	0.7 ± 0.2	0.5 ± 0.2	<0.001
S1 Dural Sac Ratio	1.1 ± 0.8	0.6 ± 0.1	<0.001
S2 Dural Sac Ratio	3.2 ± 5.5	0.6 ± 0.2	<0.001
L1 Scalloping	4.9 ± 1.6	2.8 ± 1.4	<0.001
L2 Scalloping	5.1 ± 1.6	3 ± 1.1	<0.001
L3 Scalloping	4.2 ± 1.5	2.7 ± 1.1	<0.001
L4 Scalloping	4.1 ± 1.3	2.6 ± 1.1	<0.001
L5 Scalloping	4.7 ± 3.8	2.3 ± 2.9	<0.001
S1 Scalloping	5.2 ± 2.6	2.5 ± 1.4	<0.001
S2 Scalloping	4.8 ± 2.9	2.9 ± 1.5	<0.001

Characteristics of LDS patients with pathogenic variants in the TGF-beta pathway are detailed in Supplementary data online, Table S2 (for TGFßR1 or 2 cohort) and Supplementary data online, Table S3 (for SMAD3 cohort).

MFS diagnosis performance of conventional criteria for dural ectasia

The DE criteria of the three major publications were tested on the CT-scan data of our FBN1 cohort^16–18 (see Supplementary data online, Table S4). These studies showed good sensitivity (up to 98% for Lundby et al),¹⁸ but poor specificity (best specificity of 62% for Ahn et al).¹⁶ The highest AUC reached only 0.68, indicating a poor level of performance²⁷ (see Supplementary data online, Figure S1).

Present CT-scan-based DE algorithm for MFS diagnosis

Our data showed that measurements of the antero-posterior diameter at the centre of the vertebra frequently overlapped between MFS and control patients (Table 1), indicating that variability in the DSR likely originated from the antero-posterior diameter of the spinal canal (APDSC). There was indeed a high degree of correlation between the DSR and the APDSC for each vertebra (Figure 3). The APDSC measurement being more rapidly determined than the DSR, the ease-of-use objective led us to focus on APDSC to build the models.

Figure 3.

Correlation between the DSR and the anteroposterior diameter of the spinal canal. Spearman correlation was performed between the values of the DSR and the Antero-Posterior Diameter of the Spinal Canal (APDSC) for all patients. Each dot represents the intensity of the correlation between the interconnecting variables, with the size and colour intensity indicating the r score for the correlation (0 = no correlation; 1 = perfect correlation).

To combine the dual objectives of high accuracy and ease-of-use, the number of variables was further refined, selecting the four variables with the highest importance in the RandomForest preliminary algorithm, which obtained the best sensitivity (0.84 vs. 0.74 for the GLM model) (Table 2). Hence, we focused on only three vertebrae (L1, L2 and S1), measuring the APDSC (L1 and S1) and the Scalloping (L2 and S1). Reproducibility between the two readers was good, with no significant difference (P = NS) for any of the quantitative variables. With these four measurements, algorithms with good levels of performance were obtained, particularly with the RandomForest classifier (0.84 in ROC AUC, specificity and sensitivity) (Table 2). The RandomForest-based model, named BLT-TLR, was thus selected. To limit the risk of overfitting of the model, a successful decoy test was conducted, wherein the machine failed to identify a performing algorithm when the outcome variable was random (see supplementary data online, Table S5).

Table 2.

Performances of the algorithm for MFS diagnosis

Algorithm	AUC (95% CI)	Sensitivity	Specificity	P-Value [Acc > NIR]	J index	Brier class
GLM α = 0. λ = 0.1	0.78 (0.63, 0.90)	0.74	0.84	0.000236	0.58	1.68
RandomForest	0.84 (0.69, 0.94)	0.84	0.84	1.217e-05	0.68	1.84
KNN	0.82 (0.66, 0.92)	0.79	0.84	5.808e-05	0.63	1.76
SVM	0.82 (0.66, 0.92)	0.74	0.89	5.808e-05	0.63	1.66
C5.0	0.76 (0.59, 0.89)	0.74	0.79	0.000829	0.53	1.71

Diagnostic performance of the DE algorithm vs. Ghent-1 and −2 classifications

The diagnostic performance of Ghent-1 and −2 criteria was evaluated on the MFS (FBN1) patients of the present series. Since the aim of this study was to evaluate the possibility to provide an early detection of MFS, the genetic results were not taken into consideration and the modified Ghent criteria relied on clinical and imaging parameters only. The clinical multidisciplinary evaluation was performed in the reference Centre for MFS, with the different clinicians (cardiologists, rheumatologists, ophthalmologists) presenting with a high level of expertize. Overall performance of modified Ghent scores was poor, with 67% accuracy for Ghent-1 and 62% for Ghent-2 criteria, inferior to the 94% accuracy of the new DE algorithm (see Supplementary data online, Figure S2A). However, since part of the FBN1 cohort was used for model training of the algorithm, performance was also compared on the test set only, representing 20% of the cohort. In this reduced population, accuracy was 74% for the Ghent-1 and 68% for the Ghent-2 criteria, inferior to the 84% accuracy of the DE algorithm (see Supplementary data online, Figure S2B).

Performance of the DE algorithm for LDS diagnosis

The new DE algorithm was tested on the TGFßR1&2 cohort and the SMAD3 cohort and their matched controls and obtained an AUC of 0.83 (0.73–0.90) and 0.80 (0.70–0.88), respectively, indicating a good level of performance (see Supplementary data online, Table S6). Sensitivity was fair, and specificity was excellent (93% for both cohorts).

Application for genetic HTAD screening according to the DE algorithm

To ensure ease of use for clinicians, the algorithm was used to build a Shiny app, made freely available at https://sfc-marfan.preprod.clinityx.io/. This interface offers a front page where users can enter CT-scan measurements, along with a help page detailing the methodology for obtaining them accurately. Calculation of scalloping is done automatically when entering the measures of the superior and inferior vertebral plateaus and the anteroposterior diameter at the centre of the vertebral core. The probability of MFS or LDS diagnosis (based on the results of the FBN1, TGFBR1, TGFBR2 or SMAD3 pathogenic variants cohorts) is then directly provided.

Discussion

In this study, we aimed to investigate whether parameters related to DE obtained from CT scan imaging, which is routinely performed for aortic evaluation, could be used for the identification of MFS related to FBN1 pathogenic variant. The new DE criteria identified by ML showed good performance (AUC of the ROC curve and accuracy both 0.84), superior to conventional DE criteria and even to the modified Ghent criteria, for MFS diagnosis. The DE algorithm also performed well for recognition of LDS with pathogenic variants in TGFBR1/2or SMAD3, with an AUC of 0.83 and 0.80, respectively.

DE parameters for MFS diagnosis

The contribution of DE to the diagnosis of MFS has been downgraded by the latest Ghent-2 classification, arguing that no preferred methods (CT or MRI) or uniformly accepted cut-off values for DE have emerged from the literature and that DE parameters lacked specificity.² Indeed, when using the conventional previously published DE criteria in our cohort, diagnostic performance for MFS was poor, with low specificity (see Supplementary data online, Table S4). Other studies attempted to discriminate MFS from control patients using previously published DE criteria (see Supplementary data online, Table S4):—Ahn et al, Oosterhof et al. and Lundby et al. used their own criteria, on small cohorts (maximum 44 MFS patients diagnosed according to Ghent-1 criteria with no identified mutation and 44 controls)^16–18—Weigang et al. study, involving 18 Ghent-1 MFS patients, without specified mutation, and 23 controls²⁸;—Sheikhzadeh et al study of 150 FBN1-mutated MFS patients, with no controls.²⁹ The results were highly variable, likely due to differences in study populations (old/new Ghent classification, unclear genetic status, limited samples) and variable imaging modalities.

However, given the importance of early diagnosis for proper management of MFS patients, an ML algorithm to select the best diagnostic criteria for DE was built, based on simple CT-scan imaging parameters. In an era where imaging is playing an increasingly important role, we took full advantage of whole-aorta imaging, which enabled lumbosacral cord assessment without additional acquisitions. With the development of artificial intelligence, imaging can now yield novel quantitative data and insights that were previously unavailable.³⁰ A stepwise approach was used to isolate the four most relevant parameters out of the 28 CT-scan measurements. This Random Forest-based model (BLT-TLR) showed good performance, with 84% of MFS patients with FBN1 pathogenic variant detected as such, and only 16% of age- and sex-matched controls wrongly detected as MFS. This level of performance is comparable to the published Ghent-2 criteria, including genetic results, but does not require the multiple specialists necessary for clinical assessment.¹⁵

New DE criteria for LDS detection

This study also analysed the largest population with LDS syndrome with a TGFBR1, TGFBR2 or SMAD3 variant. To date, only one patient with an identified SMAD3 variant has been included in a DE study,³¹ and the initial publication reported DE in only four out of the eight patients.¹²

Regarding TGFßR1&2 pathogenic variants, two publications with smaller samples reported a prevalence of DE slightly less important than in MFS patients but without reaching significance.^32,33 The new DE criteria performed well in LDS patients, both in patients with TGFßR1&2 (AUC 0.83) and SMAD3 pathogenic variants (AUC 0.80). Thus, our algorithm could be widely used for HTAD suspicion.

Wide applicability for MFS or LDS referral

The present diagnostic algorithm can be easily used with a dedicated application, which was made freely available online. Given the number of expert clinicians required for a proper Ghent-criteria assessment and the difficulties for patients to access these different specialists, the results may be important. The estimated time for the four measurements of the algorithm is about 5 min, and the only requirement is a good-quality CT-scan. Therefore, radiologists performing the CT-scans may be trained for these simple measurements and help in HTAD referral. Similarly, community centres without access to MFS specialists could also detect MFS or LDS patients more easily, allowing patients’ referral to specialised centres without delay.

Limitations

While promising, the present study has several limitations. First, this can be considered a monocentric study. However, patients were included in the national reference centre for MFS and related disorders, covering patients from all over France for a rare disease like MFS or LDS. In addition, management was standardised with all clinical and imaging assessments conducted by the same team. Secondly, the CT-scans were analysed retrospectively for DE parameters, with exclusion of CT-scan not allowing the lumbosacral cord assessment. This underlines the need for complete CT-scan of the entire aorta. Thirdly, controls were only matched for age and sex but not for other comorbidities. Fourthly, our algorithm does not differentiate between MFS and LDS diagnoses. It was initially designed for MFS diagnosis but also demonstrated good performance in detecting LDS. However, the main benefit of this algorithm is that it enables broader and easier disease classification, which may lead to faster referral to expert centres for diagnosis, including genetic analysis. The differential diagnosis between MFS and LDS will still require genetic analysis, but patients could benefit from earlier management. Fifth, the use of ML is associated with a high risk of model overfitting; however, the TRIPOD guidelines were followed, notably with the train/test split, and a successful decoy test was performed, which strengthens the legitimacy of our results. Finally, external validation of these results in large prospective studies are needed before implementation of the algorithm in clinical practice.

Conclusion

We developed an algorithm for new DE criteria based on four measurements of three vertebras on the largest number of patients using CT-scan imaging, demonstrating competitive diagnostic performances for MFS and LD syndromes. Moreover, a free publicly available user-friendly online interface is provided to enter the measurements. DE could therefore take on its full diagnostic importance and contribute to widespread detection of hereditary connective-tissue disorders.

Supplementary data

Supplementary data are available at European Heart Journal - Imaging Methods and Practice online.

Author contributions

Claire Bouleti (Conceptualisation, Writing—original draft, Writing—review & editing [lead], Data curation [supporting], Methodology [equal]), Raphael Thuillier (Formal analysis [lead], Methodology [equal]), Yoann Moeuf (Data curation, Investigation [lead], Writing—original draft [supporting]), Gaspard Suc (Data curation [equal], Investigation [supporting]), Laurent Gouya [Data curation, Investigation (supporting)], Florence Arnoult [Data curation, Investigation (supporting)], Noemie Tence (Data curation [equal], Investigation [supporting]), Florence Nicot (Data curation [equal], Investigation [supporting]), Benjamin Alos [Data curation, Investigation (supporting)], Margaux Gouysse (Formal analysis [lead], Methodology, Validation [equal]), Olivier Milleron [Data curation, Investigation (supporting)], Phalla Ou [Data curation, Investigation, Visualisation (equal)], Nicolas Glatt (Formal analysis [equal], Resources [supporting], Software [lead]), Maria Tchitchinadze [Data curation, Investigation (equal)], Philippe-Gabriel Steg (Methodology, Validation, Writing—review & editing [equal], Supervision [lead]), and Guillaume Jondeau (Data curation, Investigation [equal], Supervision, Writing—review & editing [supporting])

Funding

This study is supported by the French Society of Cardiology but there was no dedicated funding.

Data availability

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Lead author biography

Prof. Claire Bouleti is a cardiologist trained at Bichat Hospital, Paris, and currently working at Poitiers University Hospital since 2019. She leads the cardiac imaging department—encompassing cardiac CT, CMR, and echocardiography—as well as the Heart Failure Unit. Her main fields of interest include heart failure (particularly acute myocarditis), aortic diseases, and valvular heart diseases. Deeply involved in clinical research, she serves on the board of the French Alliance for Cardiovascular Trials (FACT network), and is responsible for research activities both nationally (French Society of Cardiology) and locally within her department.