Enhancing Chicago Classification diagnoses with functional lumen imaging probe-mechanics (FLIP-MECH)

Abstract

Background: Esophageal motility disorders can be diagnosed by either high-resolution manometry (HRM) or the functional lumen imaging probe (FLIP) but there is no systematic approach to synergize the measurements of these modalities or to improve the diagnostic metrics that have been developed to analyze them. This work aimed to devise a formal approach to bridge the gap between diagnoses inferred from HRM and FLIP measurements using deep learning and mechanics.

Methods: The "mechanical health" of the esophagus was analyzed in 740 subjects including a spectrum of motility disorder patients and normal subjects. The mechanical health was quantified through a set of parameters including wall stiffness, active relaxation, and contraction pattern. These parameters were used by a variational autoencoder to generate a parameter space called virtual disease landscape (VDL). Finally, probabilities were assigned to each point (subject) on the VDL through linear discriminant analysis (LDA), which in turn was used to compare with FLIP and HRM diagnoses.

Results: Subjects clustered into different regions of the VDL with their location relative to each other (and normal) defined by the type and severity of dysfunction. The two major categories that separated best on the VDL were subjects with normal esophagogastric junction (EGJ) opening and those with EGJ obstruction. Both HRM and FLIP diagnoses correlated well within these two groups.

Conclusion: Mechanics-based parameters effectively estimated esophageal health using FLIP measurements to position subjects in a 3-D VDL that segregated subjects in good alignment with motility diagnoses gleaned from HRM and FLIP studies.

Keywords: achalasia; dysphagia; esophageal biomechanics; esophageal motility disorders; generative AI.

Figure 1:. Mechanics-informed variational autoencoder (MI-VAE).

(A) Network architecture. The numbers above each CNN layer represent the number of channels or number of nodes in a densely connected layer. The activation parameter θ becomes input to MI-VAE to generate a latent space of 40 dimensions to which the other 6 scaler mechanics parameters are merged to define the parameter space called virtual disease landscape (VDL). (B) Workflow for developing the VDL.

Figure 2:. Examples of θ-variations for the four groups according to the FLIP-classification: normal, weak, spastic-reactive, and obstruction.

The non-dimensional time and length along FLIP are shown in the horizontal and vertical axis, respectively.

Figure 3:. Calculation of the distance matrix.

The schematic shows a 2D space for the VDL with two groups (shown in blue and red) and two data points in each group. The first step involves estimating the centroid of each group (shown by the square marker) in the VDL space by calculating the mean of each component of the data points in a group. The second step involves the calculation of the distance between each data point from the centroid of each group. The last step involves grouping these distances into the different categories and calculating the median of each category that becomes each of the elements of the distance matrix followed by scaling to add to 100 for each row.

Figure 4:. Virtual disease landscape (VDL) showing the four FLIP classification groups.

(A) VDL shown in reduced dimensions through LDA. (B) Distance matrix quantifying the effectiveness of clustering. Low values of the diagonal elements compared to the off-diagonal elements represent good separation between the groups. (C) Visualizing the contraction patterns for specific points on the VDL. The four points are selected from the extremes of the four clusters and the corresponding variations of activation parameter θ is shown.

Figure 5:. HRM diagnosis according to CCv4.0 plotted on the VDL.

(A) Reduced dimensional VDL with the points colored according to HRM diagnosis. The normal EGJ opening points are represented in blue and ones with EGJ obstruction are in red. (B) Distance matrix quantifying the effectiveness of clustering based on HRM diagnosis. The two major clusters include the one with type I, II, III achalasia and EGJOO, and the other being normal, absent contractility, IEM, and DES. (C) VDL including only EGJ obstruction cases, and (D) the corresponding distance matrix. Type I and II achalasia are overlapped on the VDL while EGJOO is distributed among the achalasia groups. (E) VDL including only normal EGJ opening cases, and (F) the corresponding distance matrix. Normal and absent contractility are the most distinct from each other while IEM is distributed uniformly among the other three groups showing great variability.

Figure 6:. Normal motility subjects according to CCv4.0 plotted on the VDL to demonstrate their variability with respect to FLIP measurements.

The colors on the VDL correspond to the maximum probability among the four FLIP-classification groups. The contraction patterns are plotted for four points on the extremes of the CCv4.0 normal motility subjects which demonstrates how the VDL separates them. Each point is associated with the probability of belonging to each of the four FLIP classification groups. The probabilities are represented by the variables PN, PW, PO, and PS corresponding to the probability of being normal, weak, obstruction, and spastic-reactive, respectively. The maximum probability is shown in red with a bold font.

Figure 7:. Subjects classified according to CCv4.0 (along rows) and their corresponding FLIP classification (along columns).

Each element refers to the number of subjects at the intersection of the two corresponding diagnosis on HRM and FLIP.

Figure 8:. Boxplots showing the probabilities of the HRM diagnosis of being one of the four groups according to FLIP classification.

The y-axis shows the probability. Type I, II achalasia and EGJOO appear mainly as obstruction on FLIP, while normal mostly appears as normal on FLIP. The other HRM-based groups appear distributed among two or more groups on FLIP. *P value < 0.05, **P value < 0.01, ***P value < 0.001. The P values not shown in the plots above correspond to several orders of magnitude less than 0.001.