Novel Lobe-based Transformer model (LobTe) to predict emphysema progression in Alpha-1 Antitrypsin Deficiency

Abstract

Emphysema, marked by irreversible lung tissue destruction, poses challenges in progression prediction due to its heterogeneity. Early detection is particularly critical for patients with Alpha-1 Antitrypsin Deficiency (AATD), a genetic disorder reducing ATT protein levels. Heterozygous carriers (PiMS and PiMZ) have variable AAT levels thus complicating their prognosis. This study introduces a novel prognostic model, the Lobe-based Transformer encoder (LobTe), designed to predict the annual change in lung density (ΔALD [g/L-yr]) using CT scans. Utilizing a global self-attention mechanism, LobTe specifically analyzes lobar tissue destruction to forecast disease progression. In parallel, we developed and compared a second model utilizing an LSTM architecture that implements a local subject-specific attention mechanism.

Our methodology was validated on a cohort of 2,019 participants from the COPDGene study. The LobTe model demonstrated a small root mean squared error (RMSE=1.73 g/L-yr) and a notable correlation coefficient (ρ = 0.61), explaining over 35% of the variability in ΔALD (R² = 0.36). Notably, it achieved a higher correlation coefficient of 0.68 for PiMZ heterozygous carriers, indicating its effectiveness in detecting early emphysema progression among smokers with mild to moderate AAT deficiency. The presented models could serve as a tool for monitoring disease progression and informing treatment strategies in carriers and subjects with AATD. Our code is available at github.com/acil-bwh/LobTe.

1. Introduction

Emphysema is characterized by abnormal, permanent enlargement of the distal airspaces due to alveolar wall destruction [1]. While the disease can occur due to various factors, Alpha-1 Antitrypsin Deficiency (AATD) is the only known genetic disorder directly associated with an increased risk for emphysema. AATD leads to low or undetectable levels of alpha-1 proteinase inhibitor (AAT), which increases the risk of early-onset, rapidly progressive lung damage [2]. Most individuals with the typical MM genotype, which represents 90%–95% of the population, express normal serum AAT levels (20–53 μmol/L). However, 5%–10% of people carry deficiency alleles, such as S or Z, leading to subnormal serum concentrations of AAT and a heightened risk of liver and/or pulmonary diseases like COPD [3]. Carriers with one normal and one mutated allele, such as PiMS and PiMZ, exhibit lower serum concentrations, ranging from 18–52 μmol/L for MS and 17–33 μmol/L for MZ. PiMZ carriers, with mild to moderate AAT deficiency, may develop disease symptoms, while the risk for PiMS carriers remains less clear. In particular, PiZZ subjects are at a significantly higher risk of developing emphysema compared with heterozygous individuals. For instance, Stoller and Aboussouan [3] showed that the risk of developing emphysema in PiZZ subjects ranges from 80% to 100%. Although the exact risk remains incompletely understood, it is evident that the majority of PiZZ individuals are highly likely to develop emphysema, therefore the prognostication tasks has a lower clinical significance. Subsequent studies, especially in PiMZ individuals, have however been inconclusive [4–7].

Forema et al. found in the COPDGene cohort that PiMZ heterozygous individuals who smoke are at increased risk for expiratory airflow obstruction and the development of emphysema compared with Z-allele non-carriers, or never-smoker PiMZ [8]. This key finding motivated us to develop a method to predict emphysema progression, specifically targeting PiMS and PiMZ heterozygous smokers, where the risk of emphysema may not be as high as in PiZZ subjects but still warrants early intervention. Investigating PiMS and PiMZ heterozygous individuals allows us to explore early radiographic markers of disease progression in a population that may not yet exhibit advanced disease but remains at risk, with significant implications for early intervention strategies. Treating these individuals promptly is crucial, given their increased risk relative to non-AATD individuals or never-smokers.

Emphysema presence and its progression can be quantified on baseline and serial CT scans through the evolution of the lung density at the 15th percentile of the adjusted lung density histogram (hereafter, lung density perc15) method (ΔALD) [9]. This measure has been extensively used as an endpoint to assess global emphysema progression [10–12] and it is commonly reported for clinical trials in people with AATD [2, 13]. Although CT scans enable the diagnosis of emphysema, they lack radiographic markers to reliably predict its evolution. Emphysema’s heterogeneity within the lung, coupled with variations in clinical presentation, physiology, therapy response, lung function decline, and survival, complicates its prognostication [14].

Building on our prior research on COPD [15], this study introduces and validates two novel deep learning models capable of capturing the complexity and spatial variability of emphysema progression. These models were designed to identify smokers with PiMS and PiMZ genotypes at higher risks of emphysema progression by predicting the change of the adjusted lung density over five years, i.e., ΔALD. Importantly, our methods are purely based on radiographic features measured at baseline and achieve results comparable to standard metrics derived from longitudinal imaging.

1.1. Related works

Convolution Neural Networks in COPD

Convolution Neural Networks (CNNs) offers a unique approach to early quantify prospective COPD outcomes by exploring the heterogeneity of the CT scan signal, improving disease prognostication in this complex population. In the last decade, CNN have made significant advances in the field of COPD due to their ability to learn highly complex representations in a data-driven manner. For example, CNNs were proposed to regress the %LAA-950 in [16]; and to regress FEV1, FEV1/FVC, and COPD severity in [17]. A long-short term memory architecture (LSTM) was combined with a CNN in [18] to classify pattern of emphysema according to absent, trace, mild, moderate, confluent, or advanced destructive (Fleischner categories, [19]) on COPDGene [20] and ECLIPSE [21] participants. Same approach was used in [22] to assess emphysema progression according to the Fleischner categories, and also fibrotic interstitial lung abnormality in [23]. CNNs were also proposed for survival prediction in patient with COPD on chest radiographs [24]. Unlike previous works, [15] proposed to use a longitudinal encoder–decoder based on convolutional layers to locally predict the emphysema progression on COPD subjects at 5-year.

Attention models in COPD

Attention mechanism have become an integral part of different deep learning approaches, for example, in LSTM the forget gate decides which information needs attention and which can be ignored, or in CNNs locally spatial and channel attention can be used to decide which information is more relevant [25]. But the transformer model was the first approach relying entirely on self-attention [26] where the attention blocks aggregate the entire input sequence. The success of the Transformer models is mainly due to the self-attention (SA) mechanism and its ability to model long-range dependencies [27]. Vision transformers (ViT) was the first major attempt to apply a pure Transformer directly to images [28]. ViT uses a vanilla transformer encoder as a replacement of the convolution operations by achieving state-of-the-art performance. However, the experimental results showed that training a vanilla transformer model on mid-sized datasets such as ImageNet (1.2 million of images), achieving accuracies of a few percentage points below ResNets of comparable size. Because transformers lack of some inductive biases inherent to CNNs such as translation and locality, they do not generalize as well as CNNs on small amount of data [29]. Therefore, the authors also suggested hybrid architecture by conjugating CNN backbone (e.g. ResNet) to Transformer.

During the last two years several transformer approaches have been proposed to better assess COPD, however, at the present, none of them were designed for prognosis. For example, [30] proposed to use a vanilla ViT approach to classify the emphysema subtypes on CT images, and [31] used a ViT approach to identify subjects with and without COPD according to FEV1/FVC < 0.7. A multimodal approach using cross modal transformer [32] was proposed in [33] to identify COPD subjects at early (GOLD 1–2) or advanced (3–4) stages of COPD according with Global Initiative for Obstructive Lung Disease (GOLD) [34]. And, [35] proposed an hybrid approach of CNNs and contextual transformer [36] to segment airway and small airway branches on CT images to assess COPD.

1.2. Contributions

While existing methods focus on COPD and emphysema progression quantification based on radiographic features [18,22–24], none forecast lung emphysema evolution for prospective emphysema progression without longitudinal imaging. In comparison with the state-of-the-art methods, four innovative aspects are presented in this work: (I) training includes learning the evolution of local emphysema progression and lung density without utilizing follow-up chest CT scans for inference; (II) a new embedding fingerprint strategy is proposed to encapsulate the local disease evolution into a unique prognostic subject signature, or five distinct prognostic lobe signatures (left/right superior lobe fingerprints, left/right inferior lobe fingerprints, and right middle lobe fingerprint); (III) the introduction and validation of two prognostic deep learning models designed to measure the emphysema progression by regressing the change of the lung density (ΔALD). The first model, named LobTe, employs a transformer encoder [26,28] on the lobe fingerprints to leverage a global self-attention mechanism for providing predictions based on lobar tissue destruction, while the second model is proposed for comparison reasons and introduces a local subject attention (LSA) based on an LSTM architecture; (IV) our study is focused on heterozygous subjects with PiMS and PiMZ genotypes at higher risks of emphysema progression;

Compared with our previous work, we are including a validation of the local emphysema progression model on heterozygous carriers (PiMS and PiMZ), a new strategy to identify prognostic fingerprints, and two prognostic models to regress the change of lung density at 5-year (ΔALD). Moreover, our models demonstrate strong association with the risk of emphysema progression on heterozygous subjects with PiMS and PiMZ genotypes.

2. Materials and methods

A total of 10,198 smokers with and without COPD across the GOLD stages were enrolled in the COPDGene (Genetic epidemiology of COPD) study [20], with more than 6000 participants returning for a second visit at 5-year (phase 2). From this population, individuals with rare deficiency alleles (e.g., FZ, MF, MI, Znull) as well as those with PiSS, PiSZ, and PiZZ genotypes were excluded. For details about the method used to identify alpha-1 genotypes and their distribution in COPDGene, please refer to [8]. A total of 4821 COPDGene participants (4391 PiMM, 288 PiMS and 142 PiMZ) with complete data and correct lobes detection were included in this study, comprising 1931 individuals across GOLD stage I-IV and 2890 without COPD, including PRISm. 2802 PiMM subjects were used for training and 2019 for testing (1589 PiMM, 288 PiMS and 142 PiMZ). Table 1 shows the distribution of genotypes for individuals with and without COPD while Table 2 shows the clinical characteristics of the populations used for training and testing.

Table 1.

Distribution of participants across genotypes and training/testing splits, stratified by COPD severity based on GOLD stage, including subjects without COPD and those with preserved ratio impaired spirometry (PRISm).

		MM			MS	MZ
		All	Train	Test	Test	Test
No COPD		2120 (48.3)	1350 (48.1)	773 (48.6)	132 (46.4)	63 (44.4)
PRISm		539 (12.3)	323 (11.5)	216 (13.6)	26 (8.9)	8 (5.6)
	1	395 (9.0)	252 (9.0)	143 (9.0)	28 (9.6)	11 (7.7)
COPD status	2	841 (19.2)	542 (19.4)	299 (18.8)	58 (19.9)	32 (22.5)
	3	403 (9.2)	282 (10.1)	121 (7.6)	41 (14.1)	19 (13.4)
	4	90 (2.1)	53 (1.9)	37 (2.3)	3 (1.0)	9 (6.3)

Table 2.

Clinical characteristics of the populations used for training and testing.

Baseline characteristics		All	Training	Test
N		4821	2802	2019
Race	nHW	3375 (70.0)	1969 (70.3)	1406 (69.6)
	AA	1446 (30.0)	833 (29.7)	613 (30.4)
Age, [y]		59.7 (8.6)	59.6 (8.7)	59.8 (8.6)
AAT genotype	MM	4391 (91.0)	2802 (100)	1589 (78.7)
	MS	288 (6.0)	-	288 (14.3)
	MZ	142 (2.9)	-	142 (7.0)
Pack-years		42.4 (23.5)	42.1 (23.9)	42.7 (23.0)
Smoking status	FS	2466 (51.2)	1423 (50.8)	1043 (51.7)
	CS	2355 (48.8)	1379 (49.2)	976 (48.3)
FEV1/FVC		0.7 (0.1)	0.7 (0.1)	0.7 (0.1)
% Emph.		5.2 (7.9)	5.2 (7.9)	5.1 (7.9)
% Emph. upper lobes		5.9 (9.2)	6.0 (9.3)	5.8 (9.2)
% Emph. lower lobes		4.1 (6.8)	4.1 (6.8)	4.1 (6.9)
No COPD		2318 (48.1)	1350 (48.2)	968 (47.9)
PRISm		573 (11.9)	323 (11.5	250 (12.4)
	1	434 (9.0)	252 (9.0)	182 (9.0)
COPD status	2	931 (19.3)	542 (19.3)	389 (19.3)
	3	463 (9.6)	282 (10.1)	181 (9.0)
	4	102 (2.1)	53 (1.9)	49 (2.4)
Change ALD [yr]		−0.2 (2.2)	−0.1 (2.3)	−0.3 (2.2)

Definition of Abbreviations: nHW: non Hispanic White; AA: Afro-American; CS: Current Smoker; FS: Former Smoker; ALD: Adjusted Lung Density. Continuous variables are presented as mean (SD) and categorical variables as N (%).

Lobe-based Transformer encoder (LobTe)

The method described in this section was designed to leverage a global self-attention mechanism for providing predictions based on lobar tissue destruction, and it consist of three main components: (1) a local density model (Fig. 1(a)) used to encode the local evolution of the lung density and emphysema progression at 5-year; (2) an embedding strategy used to summarize the local lung density embedding into five fingerprints corresponding to each lobe (Fig. 1 (b-2)); and (3) a transformer-based model (Fig. 1 (b-3)) designed to regress the evolution of emphysema based on the adjusted lung density measurements (ΔALD).

Fig. 1.

Proposed method workflow: (a) A local density model is utilized to capture the evolution of lung density and the progression of emphysema. (b) A transformer-based model is designed to predict changes in lung density (ΔALD) based on the extent of tissue destruction within each lung lobe.

Local Density Model:

The local density model introduced in [15] learns the evolution of the local density by using a longitudinal local encoder–decoder and a multilayer perceptron (MLP). This model is trained to both reconstruct a specific 32 by 32 pixel neighborhood at 5-year follow-up and predict the likelihood of emphysema progression (Fig. 1 (a) LEP). The local emphysema progression (LEP) was defined as those pixels that change from lung tissue (>−950 HU) to air (≤−950 HU) at 5-year.

Initially, the encoder–decoder is pre-trained separately for stability, followed by a combined training phase with the MLP under a conditional strategy that balances local reconstruction accuracy and emphysema progression encoding following this steps: (1) the MLP is fitted to regress the likelihood of the LEP using the encoded local patterns at 5-year; (2) the encoder is adjusted to maximize the likelihood of the LEP. In this stage the MLP model is frozen, no weight adjustments are made; and (3) the encoder–decoder is re-trained to keep a suitable local reconstruction.

The encoder incorporates a local attention mechanism and a scale mechanism via convolutional block attention modules (CBAM) [25] and dilated convolution (KSAC) [37]. A detailed description of the encoder–decoder used to learn the local density image embeddings can be seen in Table 3. The MLP is defined as a sequence of four dense blocks with 300 hidden units and one with 200 units. Each block consists of a dense layer, a batch normalization and a dropout strategy with a ratio of 40%. A swish function is used as activation for all the layers with the exception of the latent space z and the reconstruction output and LEP regression. After each convolution, including those in KSAC and CBAM, a batch normalization strategy is applied.

Table 3.

Autoencoder architecture used in the local density model.

Encoder				Decoder
Op.	Kernel	Stride	Output	Op.	Kernel	Stride	Output
KSAC			32 × 32 × 128	Dense + BN			2048
CBAM	3	2	16 × 16 × 128	Reshape			4 × 4 × 128
KSAC			16 × 16 × 128	Conv + BN	3	1	4 × 4 × 128
CBAM	3	2	8 × 8 × 128	Up-Conv	3	2	10 × 10 × 128
CBAM	3	1	8 × 8 × 128	Conv + BN	3	1	8 × 8 × 128
CBAM	3	2	4 × 4 × 128	Up-Conv	3	2	18 × 18 × 128
CBAM	3	1	4 × 4 × 128	Conv + BN	3	1	16 × 16 × 128
Flatten			2048	Up-Conv	3	2	32 × 32 × 128
Dense (z)			300	Conv	3	1	32 × 32 × 1

KSAC: Kernel Sharing with rate = [2, 4]; CBAM: Convolutional Block Attention; Up-Conv: Upsampling, Padding, Conv (kernel size = 1) and Batch normalization (BN).

The loss used to train the longitudinal encoder–decoder is defined as follow Curiale and Estépar [15]:

$$\begin{array}{l}{L}_{\mathrm{A}\mathrm{E}}=\frac{1}{N}\sum _i{\lambda }_0{L}_{r\Vert \mathrm{l}\mathrm{e}\mathrm{p}}({\mathbf{y}}^{\left(i\right)},{\stackrel{\mathbf{ˆ}}{\mathbf{y}}}^{\left(i\right)})+{\lambda }_1{L}_{r\Vert \mathrm{e}}({\mathbf{y}}^{\left(i\right)},{\stackrel{\mathbf{ˆ}}{\mathbf{y}}}^{\left(i\right)})\\ +{\lambda }_2{L}_r({\mathbf{y}}^{\left(i\right)},{\stackrel{\mathbf{ˆ}}{\mathbf{y}}}^{\left(i\right)})+{\lambda }_3{L}_{r\Vert \mathrm{l}\mathrm{u}\mathrm{g}\mathrm{g}}({\mathbf{y}}^{\left(i\right)},{\stackrel{\mathbf{ˆ}}{\mathbf{y}}}^{\left(i\right)})\\ +{\lambda }_4{L}_{r\Vert \mathrm{e}\mathrm{p}\mathrm{r}}({\mathbf{y}}^{\left(i\right)},{\stackrel{\mathbf{ˆ}}{\mathbf{y}}}^{\left(i\right)})+\sum _s{\lambda }_5{S}_w\left(s\right)L_{r\Vert s}({\mathbf{y}}^{\left(i\right)},{\stackrel{\mathbf{ˆ}}{\mathbf{y}}}^{\left(i\right)}),\end{array}$$ (1)

where ${\mathbf{y}}^{\left(i\right)}$ corresponds to a patch of 32 by 32 pixels in a 5-years co-register follow-up scan, and ${\stackrel{\mathbf{ˆ}}{\mathbf{y}}}^{\left(i\right)}$ its predicted reconstruction. $L_r$ is the mean absolute error in the patch

$$L_r\left(\mathbf{y},\stackrel{\mathbf{ˆ}}{\mathbf{y}}\right)=\sum _j|y_j-{\hat{y}}_j|,$$ (2)

and $L_{r\Vert \mathrm{m}\mathrm{a}\mathrm{s}\mathrm{k}}(\mathbf{y},\stackrel{\mathbf{ˆ}}{\mathbf{y}})$ is the mean absolute error conditioned to those pixels in the patch where mask is true, i.e. $\mathrm{m}\mathrm{a}\mathrm{s}\mathrm{k}(i,j)=\mathrm{T}\mathrm{r}\mathrm{u}\mathrm{e}$. In Eq. (1) mask corresponds to those pixels with emphysema progression ($lep$), emphysema measured at 5-years follow-up ($e$), emphysema prediction ($epr$), lung tissue ($lung$) or pixels with a particular emphysema subtype ($s$). Moreover, the emphysema subtype [14] is used to acknowledge that different subtypes have different patterns of progression based on their severity. In this way, $S_w$ is defined as follow: 0.4,0.8,1,0.3,0.2 and 0.2 for paraseptal emphysema, mild, moderate and severe centrilobular emphysema, panlobular emphysema and normal parenchyma respectively. The LEP loss function, i.e., the loss used to regress the likelihood of the emphysema progression, was defined as the mean absolute error.

Embedding fingerprint:

The local model encoder is applied on $N$ local patches for each lobe extracted with stride of 4-pixels over the lobe region to derive $N\times 300$ embeddings that are aggregated into a fingerprint using the deciles of the embedding distribution resulting in a representation of 300 by 11 embedding percentiles (0%, 10%, … 100%) per lung lobe as it can be seen in Fig. 1 (b-2).

Subject Density Model:

A new prognostic models (LobTe) is proposed to regress the emphysema progression according to the change of the lung density. This model incorporates a self-attention mechanisms based on a pure transformer encoder, substituting traditional ViT image patches with lobe embedding fingerprints to maintain global attention according to lobes’ positions (see Fig. 1 (b-3)). This adaptation reduces the model size to under 141,631 parameters, addressing the challenge of ViT models’ reliance on large datasets [28]. The model consists of one layer of 8 multi-head self attention blocks with a dropout ratio of 25%, a MLP of 32 units with a GELU non-linearity, and an embedding representation of the global deep-phenotype of size 32. The outputs correspond to a dense layer of 1 unit with a linear activation. The input of the model is normalized according to the z-score measured on the training dataset ($\mu =0.0186$ and $\sigma =2.868$).

Local Subject Attention model (LSA)

For the purpose of comparison, we propose a slightly different approach to predict the change of lung density. This model, named LSA, is based on an LSTM architecture providing a local subject attention mechanism. Like the transformer-based model, the LSA model is built on top of the Local Density Model (see 1(a)). However, instead of feeding the model with the lobe fingerprints, the LSA model receives a unique subject fingerprint of 300 by 11 embedding percentiles as it was described in the section Embedding fingerprint (Fig. 2).

Fig. 2.

Local Subject Attention model (LSAM) proposed to regress the change of lung density (ΔALD) based on a unique subject fingerprint.

The architecture proposed for the LSA model comprises two layers: an LSTM with 10 neurons, where the deep-phenotype percentile dimension is interpreted as a spatial-time dimension; and a dense layer with 5 neurons employing a batch normalization strategy. The LSTM layer uses a hyperbolic tangent as activation while the dense layer uses a swish function. Similar to the LobTe model, the output correspond to a dense layer of 1 unit with a linear activation and the input is normalized according to the z-score described before, i.e., $\mu =0.0186$ and $\sigma =2.868$.

Training

The local density model (Fig. 1(a)) was trained on 6.8 million random co-registered patches extracted from the baseline and follow-up CT in 984 COPDGene participants randomly selected from the training set. The LSTM and LobTe models (Fig. 1 (b-3) and (c-3)) were fitted on 2802 COPDGene participants reserved for training. The models were trained using a 5-folding cross validation strategy for early stopping and boosting their performance. The model ensemble used for inference corresponds to the average output of these 5 models predictions. The population reserved for training was randomly selected from our cohort to preserve as much as possible the GOLD distribution observed in the COPDGene cohort. The local density model was optimized as it was described in [15] while the LSTM and LobTe models were optimized using a stochastic gradient descent (Adaptive Moment Estimation) with a learning schedule defined as follow:

$$\mathrm{l}\mathrm{r}=\frac{1}{\sqrt{G}}\mathrm{m}\mathrm{i}\mathrm{n}\left(\frac{1}{\sqrt{i}},i\frac{1}{\sqrt{{\mathrm{w}}^3}}\right)$$

where $\mathrm{G}$ corresponds to the dimension of the global lobe fingerprint ($\mathrm{G}=32$) and $i$ is the iteration step. The warm-up factor, w, was set to

$$\mathrm{w}=⌊\frac{N_e}{4}⌋⌊\frac{N}{\mathrm{b}\mathrm{s}}⌋,$$

with the maximum number of epochs, $N_e=1000$, the number of training samples, $\mathrm{N}=2802$, and the batch size used, $\mathrm{b}\mathrm{s}=32$.

3. Results

We began our analysis by studying the performance of the local density model to identify lung regions at risk of emphysema progression on our testing population with heterozygous carriers (291 PiMS and 142 PiMZ) and 251 testing subjects without AATD (PiMM). A region is considered at risk of emphysema progression (EP) when the percentage of pixels with EP is greater than 5% as it was previously defined in [15].

A total of 13.850 million of non-overlapping regions (~20,000 regions per subject) with a prevalence of regions at EP risk of 17.7% for PiMM, 18.1% for PiMS, and 23.5% for PiMZ were included in this analysis. The local density model shows a good performance in identifying lung regions at risk of EP for subjects with normal (PiMM) and mutated AAT genes (PiMS and PiMZ heterozygous carriers). Fig. 3 shows the roc and precision–recall curves stratified according to PiMM, and PiMS and PiMZ heterozygous carriers. The area under the roc curves (ROC-AUC) measured for the local density model are 87.3%, 87.5% and 86.9% for PiMM, PiMS and PiMZ respectively, and the area under the precision–recall curves (PR-AUC) are 53.9% for PiMM, 53.4% for PiMS and 60.2% for PiMZ. Unlike the ROC curve, which is invariant to imbalanced datasets, the PR curve of a random classifier would appear as a horizontal line, with precision proportional to the number of positive examples in the dataset. In our case, the PR-AUC for a random classifier would be 17%, 18.1%, and 23.5% for PiMM, PiMS, and PiMZ, respectively.

Fig. 3.

Performance of the local density model, including ROC and precision–recall curves, stratified for heterozygous carriers (PiMS and PiMZ) and individuals without AATD in the testing dataset.

Table 4 summarizes the performance of the local density model to identify regions at risk of EP at 5-year and the 95% confidence interval (CI) measured on 100 bootstrap samples. At an operating point that maximizes the Youden’s index on the training dataset (probability Th = 0.158), the model’s sensitivity (Sens.), specificity (Spec.), F1-score, and Cohen’s Kappa coefficient (K) were Sens. = 54%, Spec. = 90.9%, F1 = 54.9%, and K = 45.5% for non AATD subjects, and Sens. = 55.5% and 60%, Spec. = 90.2% and 88.7%, F1 = 55.6% and 61%, K = 45.8% and 49.2% for heterozygous subjects with PiMS and PiMZ genotypes respectively. The results generally show a good precision to identify regions at risk of EP with a moderate agreement (K ∈ [41%, 60%]), especially for heterozygous carriers (PiMS and PiMZ).

Table 4.

Local density model performance for predicting lung regions at risk of emphysema progression on non-AATD and heterozygous (PiMS and PiMZ) subjects.

	PiMM [%]	PiMS [%]	PiMZ [%]
Prev.	17.7 (17.6, 17.8)	18.1 (18.0, 18.1)	23.5 (23.4, 23.5)
Sens.	54.0 (53.9, 54.0)	55.5 (55.4, 55.6)	60.0 (59.9, 60.0)
Spec.	90.9 (90.8, 90.9)	90.2 (90.2, 90.2)	88.7 (88.6, 88.7)
PPV	55.9 (55.8, 56.0)	55.7 (55.6, 55.7)	62.0 (61.8, 62.0)
FPR	9.1 (9.10 9.14)	9.8 (9.7, 9.8)	11.3 (11.2, 11.3)
FNR	46.0 (45.9, 46.1)	44.5 (44.3, 44.5)	40.0 (39.9, 40.0)
ACC	84.4 (84.3, 84.4)	84.0 (83.9, 84.0)	81.9 (81.9, 81.9)
BA	72.4 (72.3, 72.5)	72.9 (72.8, 72.9)	74.3 (74.2, 74.3)
Fl	54.9 (54.8, 55.0	55.6 (55.5, 55.6)	61.0 (60.9, 61.0)
K	45.5 (45.3, 45.5)	45.8 (45.7, 45.8)	49.2 (49.1, 49.3)
ROC-AUC	87.3	87.5	86.9
PR-AUC	53.9	53.4	60.2

Sens.: sensitivity; Spec.: specificity; PPV: positive predictive value; FPR: false positive rate (FPR); FNR: false negative rate; ACC: accuracy; BA: balanced accuracy; F1-score; K: Cohen’s Kappa coefficient (K); PR: Precision-Recall; ROC: area under the curves. All the values, including the 95% CI, are expressed in percentage.

Prediction of the change of lung density (ΔALD) on heterozygous (PiMS and PiMZ) individuals and non-AATD

Results on our testing dataset (n = 2019) show small root mean squared errors (RMSE) and good correlation coefficients (ρ) between the model proposed to predict ΔALD, i.e., LobTe prediction, and the measurement at 5-year (RMSE = 1.73 g/L-yr and ρ = 0.61). The LobTe model successfully explain more than 35% of the ΔALD variability with coefficient of determination R² = 0.36. Fig. 4 shows the scatter plots, including a linear fit and the Pearson’s correlation coefficient (ρ), and Bland–Altman plot between the model’s prediction and the observed annualized ΔALD [g/L-yr]. The Bland–Altman plots show good agreement with small bias and standard deviation errors (μ = 0.09 and σ = 1.73 g/L-yr).

Fig. 4.

Performance of the proposed method (LobTe) to predict the annualized change in lung density ΔALD [g/L-yr]: Scatter plot with a linear fit and Pearson’s correlation coefficient, along with Bland–Altman plots comparing the LobTe predictions to the observed annualized changes.

Additionally, the model show significant predictive power to regress ΔALD on heterozygous carriers, for example, the model shows small errors (PiMS_RMSE = 1.66 g/L-yr and PiMZ_RMSE = 1.73 g/L-yr), and good Pearson’s correlation (PiMS_ρ = 0.58 and PiMZ_ρ = 0.68) as it can be seen in Fig. 5. Smoking history is an important risk factor for the development of emphysema, particularly in PiMZ heterozygous individuals. Table 5 presents the model’s performance and the prevalence for PiMM, PiMS, and PiMZ, stratified by smoking history over a 5-year follow-up period. Our findings indicate a decrease in the model’s performance for PiMM and PiMS between former and current smokers. On the other hand, we observed an improvement in the model’s performance for PiMZ, with the Pearson correlation coefficient increasing from 0.68 in former smokers to 0.73 in current smokers. However, due to the small prevalence of individuals, particularly those who transitioned between smoking statuses (less than 3%), further validation of these results is necessary before drawing definitive conclusions.

Fig. 5.

Performance of the proposed method (LobTe) in predicting the annualized change in lung density (ΔALD [g/L-yr]) for PiMS and PiMZ heterozygous smokers, as well as non-AATD (PiMM) smokers: This includes scatter plots with a linear fit and Pearson’s correlation coefficient, as well as Bland–Altman plots comparing the LobTe predictions to the observed annualized ΔALD [g/L-yr].

Table 5.

Pearson’s correlation coefficient and prevalence for PiMM, PiMS and PiMZ subjects stratified by smoking history during the 5-year follow-up.

	ρ			Prevalence (n = 2019)
	PiMM	PiMS	PiMZ	PiMM	PiMS	PiMZ
F	0.632	0.618	0.678	36.2%	8.7%	4.7%
C	0.574	0.551	0.730	31%	4.1%	1.6%
C → F	0.620	0.530	0.662	9.8%	1.2%	0.4%
F → C	0.781	0.414	0.548	1.6%	0.4%	0.2%

F: Former smoker; C: Current smoker; C → F: Current smoker in baseline that transitioned to former smoker during the 5-year follow-up; F → C: Former smokers that transitioned to current smoker during the 5-year follow-up.

The error distribution made by the LobTe prediction and the observed annualized ΔALD [g/L-yr] on heterozygous subjects (PiMS and PiMZ) and non-AATD (PiMM) stratified by Pre-COPD (GOLD 0 and PRISm) and COPD (GOLD I-IV) is presented in Fig. 6. This analysis shows no significant differences between the error of predicting the change of ALD for heterozygous subjects, showing that the LobTe prediction could be a useful tool to predict the evolution of ΔALD on subjects with mutated AAT genes on an early stage of COPD.

Fig. 6.

Error distribution between the LobTe prediction and the observed annualized ΔALD [g/L-yr] on heterozygous subjects (PiMS and PiMZ) and non-AATD (PiMM). Pre-COPD: GOLD 0 and PRISm; COPD: GOLD I-IV.

Comparison between lobar and subject attention mechanisms (LobTe vs. LSA)

Next, we compare the prediction of the LobTe and LSA models to regress ΔALD at 5-year. The scatter plots for the LSA model, including a linear fit and Bland–Altman plot between the model’s prediction and the observed ΔALD g/L-yr, can be seen in Fig. 7. Both models have shown significant predictive power to regress ΔALD with similar performance in terms of error and RMSE (1.76 g/L-yr for the LSA model). However, the lobe-based transformer model shows a slight improvement of 3% on the correlation and R² (ρ = 0.58 and R² = 0.33 for the LSA model).

Fig. 7.

Performance of the local subject attention (LSA) model to predict the annualized change in lung density ΔALD [g/L-yr]: Scatter plots with a linear fit and Pearson’s correlation coefficient, along with Bland–Altman plot comparing the LSA model predictions to the observed annualized changes.

Prediction of high-risks EP on heterozygous (PiMS and PiMZ) individuals and non-AATD

Lastly, we studied performance of our model, LobTe, to predict heterozygous subjects (PiMS and PiMZ) at high-risk of emphysema progression (EP) according to distinct progression thresholds defined as ΔALD < 0 g/L-yr and the 55th (ΔALD < −0.459 g/L-yr), 65th (ΔALD < −0.963 g/L-yr) and 75th (ΔALD < −1.317 g/L-yr) percentiles of the adjusted lung density evolution of non-smokers controls¹ measured at 5-year (n = 68). The model’s performance on heterozygous subjects (PiMS and PiMZ) and non-AATD (PiMM) was analyzed for each group by means of the sensitivity (Sens.), specificity (Spec.), positive predictive value (PPV), false rates (positive (FPR) and negative (FNR)), accuracy (ACC), balanced accuracy (BA), F1-score and Cohen’s Kappa coefficient (K). Additionally, the 95% confidence interval was measured on 100 bootstrap samples. For the sake of brevity, we will omit the same analysis on the LSA model.

The model’s performance in detecting individuals at risk of EP, both in heterozygous carriers and non-AATD smokers, is summarized in Table 6. Across all risk groups, the LobTe model achieved good accuracy and F1 scores, particularly in heterozygous subjects, with mean values of ACC = 72.92%, Balanced ACC = 69.79%, and F1 = 64.25%, alongside relatively low false rates (mean FPR = 23.51% and FNR = 36.89%). Cohen’s Kappa coefficient indicated moderate agreement (mean K = 40.13%), with the highest agreement observed in PiMZ subjects (mean K = 47.2%), specially for subjects with lung density decrement lower than 1 g/L-yr (K = 0.57.1, K = 50.9 and K = 45.4). In contrast, the group of non-AATD smokers demonstrated lesser agreement (mean Kappa K = 33.75%). Notably, agreement decreased by approximately 10.16% (mean K = 28.33%) among subject groups with a lung tissue loss below 1.317 g/L-yr, predominantly comprising non-AATD smokers participants (K∈[39.5, 26.8]). Heterozygous subjects at mild (GOLD 1–2) and severe (GOLD 3–4) stages of COPD exhibited median (mean) ΔALD values of 0.39% (0.87%) and 2.76% (3.4%), respectively.

Table 6.

LobTe performance for predicting non-AATD and heterozygous (PiMS and PiMZ) subjects at risk of emphysema progression.

Population		High-risk of EP < ΔALD [g/L-yr] (95% CI)
		<0 [g/L-yr]	<-0.459 [g/L-yr]	<-0.963 [g/L-yr]	<-1.317 [g/L-yr]
PiMM (n = 1589)	Prev.	59.3 (57.5, 61.9)	49.4 (47.6, 50.9)	38.8 (37.1, 40.7)	31.8 (29.9, 33.9)
	Sens.	80.7 (78.5, 82.2)	72.4 (70.5, 74.9)	54.9 (51.4, 58.1)	39.7 (36.3, 42.8)
	Spec.	57.9 (54.8, 60.4)	63.8 (60.4, 65.9)	77.2 (75.5, 79.3)	85.0 (83.3, 86.4)
	PPV	73.7 (71.4, 75.9)	66.1 (63.8, 67.9)	60.4 (57.2, 63.4)	55.4 (51.5, 59.3)
	FPR	42.1 (39.6, 45.2)	36.2 (34.1, 39.6)	22.8 (20.7, 24.5)	15.0 (13.6, 16.7)
	FNR	19.3 (17.8, 21.5)	27.6 (25.1, 29.5)	45.1 (41.9, 48.6)	60.3 (57.2, 63.7)
	ACC	71.4 (69.7, 72.9)	68.0 (66.3, 69.4)	68.5 (66.8, 70.5)	70.6 (68.7, 72.3)
	BA	69.3 (67.5, 70.9)	68.1 (66.4, 69.4)	66.1 (64.3, 67.7)	62.4 (60.4, 64.2)
	F1	77.0 (75.4, 78.4)	69.1 (67.4, 70.9)	57.6 (55.0, 60.1)	46.3 (43.0, 49.2)
	K	39.5 (35.8, 42.7)	36.1 (32.7, 38.8)	32.6 (29.2, 36.2)	26.8 (22.6, 30.6)
PiMS (n = 288)	Prev.	59.4 (54.5, 62.9)	50.7 (45.8, 55.6)	36.8 (31.6, 41.3)	27.8 (24.0, 31.6)
	Sens.	77.8 (72.1, 83.2)	69.9 (63.1, 76.0)	52.8 (46.0, 60.8)	37.5 (30.1, 48.6)
	Spec.	68.4 (61.1, 75.2)	66.9 (61.3, 73.5)	73.6 (67.0, 78.4)	83.7 (78.8, 88.1)
	PPV	78.2 (73.2, 83.4)	68.5 (62.8, 75.4)	53.8 (46.8, 61.8)	46.9 (36.6, 58.6)
	FPR	31.6 (24.8, 38.9)	33.1 (26.5, 38.7)	26.4 (21.6, 33.0)	16.3 (11.9, 21.2)
	FNR	22.2 (16.8, 27.9)	30.1 (24.0, 36.9)	47.2 (39.2, 54.0)	62.5 (51.4, 69.9)
	ACC	74.0 (69.8, 78.1)	68.4 (63.2, 72.9)	66.0 (62.1, 70.5)	70.8 (66.6, 75.7)
	BA	73.1 (69.0, 76.9)	68.4 (63.2, 73.0)	63.2 (59.1, 67.4)	60.6 (55.9, 67.6)
	F1	78.0 (73.6, 81.9)	69.2 (63.2, 74.5)	53.3 (48.3, 59.2)	41.7 (34.9, 52.2)
	K	46.1 (37.5, 54.1)	36.8 (26.4, 45.8)	26.6 (18.5, 35.1)	22.5 (12.1, 36.5)
PiMZ (n = 142)	Prev.	56.3 (50.0, 62.7)	43.7 (38.0, 50.0)	37.3 (31.0, 43.0)	30.3 (23.9, 37.4)
	Sens.	81.2 (73.3, 88.8)	80.6 (70.3, 87.7)	58.5 (47.9, 68.7)	46.5 (35.8, 58.3)
	Spec.	75.8 (66.7, 84.1)	71.2 (62.2, 80.0)	85.4 (78.5, 91.8)	86.9 (82.3, 91.7)
	PPV	81.2 (73.3, 87.4)	68.5 (58.7, 76.3)	70.5 (59.0, 81.3)	60.6 (48.2, 72.7)
	FPR	24.2 (15.9, 33.3)	28.7 (20.0, 37.8)	14.6 (8.2, 21.5)	13.1 (8.3, 17.7)
	FNR	18.8 (11.2, 26.7)	19.4 (12.3, 29.7)	41.5 (31.3, 52.1)	53.5 (41.7, 64.2)
	ACC	78.9 (73.2, 83.8)	75.4 (68.3, 79.6)	75.4 (68.3, 81.7)	74.6 (69.0, 79.6)
	BA	78.5 (72.9, 83.7)	75.9 (69.8, 80.9)	71.9 (64.5, 77.8)	66.7 (60.7, 72.9)
	F1	81.2 (75.6, 86.4)	74.1 (65.7, 79.2)	63.9 (55.4, 72.3)	52.6 (41.1, 61.2)
	K	57.1 (45.4, 67.3)	50.9 (37.9, 59.4)	45.4 (29.8, 56.7)	35.7 (23.0, 47.0)

Sens.: sensitivity; Spec.: specificity; PPV: positive predictive value; FPR: false positive rate (FPR); FNR: false negative rate; ACC: accuracy; BA: balanced accuracy; F1-score; K: Cohen’s Kappa coefficient (K). All the values, including the 95% CI, are expressed in percentage.

4. Discussion and conclusions

Several studies have demonstrated that individuals exhibiting similar declines in lung density (ΔALD < 0) across different lung locations can experience diverse degrees of airflow obstruction [38]. It has been established that declines in lobe-specific lung density are indicative of emphysema progression. Capitalizing on this insight, our study introduces and validates a novel prognostic model based on a transformer architecture, designed to predict the annual change in lung density (ΔALD [g/L-yr]). This model employs a global self-attention mechanism to analyze lobar tissue destruction.

Additionally, for comparative purposes, we developed a second prognostic model named LSA, which utilizes an LSTM architecture to implement a local subject-specific attention mechanism. Although both models demonstrate robust predictive capabilities for ΔALD, the Lobe-Transformer model (LobTe) surpasses the LSTM-based model (LSA) by showing a 3% improvement in correlation and R² values. We hypothesize that the performance gap between the models may widen with increased dataset sizes, a typical scalability advantage of transformer-based models.

The capability of our model in predicting changes in lung density from baseline CT scans underscores their potential utility for the early detection of heterozygous carriers (PiMS and PiMZ) at elevated risk of emphysema progression, particularly in subjects with mild to moderate AAT deficiency (PiMZ), where the LobTe model achieved a correlation coefficient of 0.68. While a 35% explained variability may seem moderate, in the context of complex diseases like emphysema, this represents a significant contribution. EP is influenced by a multitude of factors, and capturing 35% of the variability provides valuable predictive power that can inform clinical decisions. For example, current state-of-the-art clinical-based methods explain less than 15% of the variance in this context [39]. As no never-smoking participants were included in this study, there is no data available for the PiMS and PiMZ populations as a whole, which represents the main limitation of this work. However, we would expect the model’s predictive power for never-smokers with deficiency alleles (S or Z) to be similar to that of individuals with the normal genotype (PiMM) who do not have COPD.

It is important to note that CT scans are a cost-effective, accessible, and practical for routine clinical use. The primary clinical rationale for developing this prediction model is to facilitate early identification of individuals at higher risk of EP progression, enabling timely interventions. Even slight reductions in AAT levels, when combined with other risk factors, can have a significant impact on disease progression. The model could provide a personalized risk assessment, which is more informative than considering AAT levels or genotype alone. While smoking cessation is universally recommended, patients identified as high-risk might benefit from more intensive interventions, such as personalized counseling, closer monitoring, or consideration of pharmacotherapy to assist in smoking cessation. Additionally, informing patients of their personalized risk can enhance their motivation to quit smoking and adhere to treatment plans. To our knowledge, no existing CT-based methodologies in the domain of lung diseases, including Chronic Obstructive Pulmonary Disease (COPD) and AATD, currently predict changes in emphysema based on adjusted lung density (Perc15). Our findings are validated with data from an observational longitudinal study, providing a solid foundation for future research aimed at anticipating changes in lung density.

Results from our study indicate that the proposed method can predict changes in ALD over one year with notable accuracy and consistency across varying degrees of disease severity and progression rates. Furthermore, the discrepancies between predicted and observed annualized ΔALD for pre-COPD and COPD heterozygous carriers (PiMS and PiMZ), as well as non-AATD smokers (PiMM), were minimal and not statistically significant. However, we observed diminished performance of the LobTe model in predicting high-risk emphysema progression among heterozygous subjects (PiMS and PiMZ) and non-AATD smokers when lung tissue loss was below 1.317 g/L-yr, predominantly affecting PiMS and non-AATD smoker participants.

In conclusion, the presented model could serve as a tool for monitoring disease progression and informing treatment strategies in AATD patients. Moreover, our study paves the way for further investigations, underscoring the necessity of validation across larger and more diverse cohorts. Such efforts will not only corroborate our findings but also foster personalized management and treatment strategies for COPD and AATD, potentially enhancing the allocation of healthcare resources.