Binary Classification of Lung Cancer Subtypes Using Hybrid CNN–BiLSTM and Attention-Based Dense GRU Models

Lung cancer is still one of the most common and lethal cancers, representing a considerable portion of cancer‑related deaths. Non‑Small Cell Lung Cancer (NSCLC) and Small Cell Lung Cancer (SCLC) account for the major clinical classifications of lung cancer with distinct pathological, genomic, and treatment profiles [1]. Distinguishing between NSCLC and SCLC is important as there are meaningful differences in treatment approaches and prognoses [2]. Although the traditional types of diagnosis—histopathology and imaging have been reliable for diagnoses, both are limited in their capability to differentiate early and to scale, particularly when limited resources are available [3]. Emerging research in machine learning and deep learning provides new opportunities to automate the classification of cancer subtypes using demographic and clinical data [4, 5]. In this study, we aim to classify lung cancer into NSCLC and SCLC as a supervised learning problem with structured tabular data. We framed the task as a binary classification problem, with the intention of building intelligent, data‑driven diagnostic tools that will help increase early detection, clinical decision‑making support, and personalized treatment planning [6]. Using a public Kaggle dataset, we present and assess new deep learning models for distinguishing between these two important lung cancer subtypes.

Various deep learning and machine learning models have been used for lung cancer subtype classification, including convolutional neural network (CNN) architectures, multi‑layer perceptron’s (MLPs), and XGBoost‑based classifiers [7]. Despite their good performance on image data, CNNs frequently perform poorly on structured tabular datasets where they cannot assess the importance of features dynamically [8]. MLPs are fairly simple and widely used, but do not flexibly focus on feature selection and often lead to poor generalization on noisy or imbalanced clinical data. XGBoost, along with other ensemble methods, are powerful predictors for binary, categorical, and multiclass tasks, but they result in black‑box models that lack interpretability, an important requirement for healthcare use. Furthermore, when mixed‑type features (numerical/categorical) are used as supervised variables, such weaknesses undermine trust and reduce clinical usability.

To address these difficulties, there is an urgent need for deep learning architectures that guarantee not only high classification accuracy but also transparency, feature awareness, and adaptability with real clinical data. This leads us to investigate adopting the TabNet and FT‑Transformer architectures that are designed to leverage tabular data, accommodate mixed features, and provide inherent transparency through attention and feature masks [9]. We developed two specialized models for NSCLC vs. SCLC classification, which we expect to improve classification accuracy and confidence while providing a more intuitive framework for explainable decision‑making within a clinical workflow. This step will contribute to the larger goal of developing intelligent, interpretable, and trustworthy AI tools for the diagnosis and subtype prediction of cancer [10].

1.1 Key objectives

The following is the key contribution of the research:

• To ensure reliable model training, preprocessing was performed by handling missing values through imputation and encoding categorical variables. To standardize the feature space, numerical attributes were normalized using standard scaling for consistent model performance.
• To propose a novel hybrid architecture (Att-DGRU) that integrates attention mechanisms with Bidirectional GRUs and dense layers for effective contextual learning in SCLC classification.
• To demonstrate Non-Small Cell Lung Cancer (NSCLC) classification using a novel hybrid approach called hybrid CNN–BiLSTM.

Moreover, the organization of the study is structured as follows: Section 1 provides the introduction of the study, recent literatures related to lung cancer classification is discussed in Section 2, and the proposed model has been discussed in Section 3. Moreover, the result and discussion are given in Section 4, and the conclusion for the study has been given in Section 5.

The workflow diagram shown in Figure 1 is a detailed framework for lung cancer classification using a two-stage hybrid deep learning approach. The pipeline begins with input (the data to be classified, whether radiomic or clinical features), and a pre-processing module, where we included all of the requisite tasks, like data cleaning, encoding, and normalising the features of the dataset (to maximise learning).

Figure 1. Architecture of the proposed model

Subsequently, the pre-processed dataset is used in the CNN–BiLSTM model that classifies Non-Small Cell Lung Cancer (NSCLC) by inferring spatial and temporal patterns. Following classification of NSCLC, the refined Att- DGRU model is used to classify 2Small Cell Lung Cancer (SCLC) using sequential dependencies and relevant features produced by the Att-DGRU model. In the final stage of the framework, we evaluated the performance of our classification method based on accuracy, precision, recall and F1-score. Without detrimentally affecting reliability, the layered approach we present here should allow for automated and accurate diagnostic classification in lung cancer screening.

3.1 Dataset description

The Lung Cancer Data in Kaggle, hosted by Andrew Mvd, is available with patient-level information that can be used to categorize the type of lung cancer. This dataset consists of around 300 records that belong to two main classes: Non-Small Cell Lung Cancer (NSCLC) and Small Cell Lung Cancer (SCLC). It is offered in CSV (table) format, with the primary clinical items being Age, Gender, Smoking Status, Symptoms, and the type of Cancer (label) he/she has.

In this paper, a feature engineering was applied to the dataset to enhance clinical interpretability and model performance. Other derived attributes are:

Air Pollution Exposure- an environmental score that represents the long-term exposure to pollutants.

Genetic Risk Index- the summative gene of hereditary vulnerability determined by reported family history and genetic markers.

Symptom Score- composite score obtained by summing up several symptom indicators (e.g., cough, chest pain, weight loss, fatigue).

Other causes- alcohol consumption, job risk, and chronic illness.

Moreover, the dataset link is given as follows, https://www.kaggle.com/datasets/andrewmvd/lung-cancer-dataset and the following Table 2 shows the key features in the dataset.

Table 2. Key features in the dataset

3.2 Preprocessing

Proper preprocessing is crucial for ensuring data integrity, consistency, and usability in training deep learning models. This study included three main preprocessing components: cleaning, encoding, and feature scaling. The focus was on Standard Scaler, a prevalent normalization technique for numerical data [17].

Data Cleaning: Missing or null values in the dataset were managed using:

Row removal for records with substantial (> 90%) null values.

•   Mean imputation was used for numerical fields and mode imputation was used in categorical fields.

Let $D=\{x 1, x 2, \ldots$,} denote the dataset. If a value $x i$ is missing, then use mean imputation as given in Eq. (1),

$x_i=\frac{1}{N} \sum_{j=1}^N \frac{x}{j}$               (1)

Label Encoding: Convert binary target class into numerical form using direct mapping and the binary target class, cancer type, was encoded using Eq. (2),

$N S C L C \rightarrow 0, S C L C \rightarrow 1$               (2)

Categorical Feature Encoding: Nominal features such as Gender and Smoking Status were encoded with One- Hot Encoding creating separate binary columns for each class. Let $C \in\{c 1, c 2, \ldots$,} be a categorical variable. Then it can be mathematically deliberated using Eq. (3),

$v_i=\left\{\begin{array}{l}1 \\ 0\end{array}\right.$ if $c=c_i$              (3)

Feature Scaling (Standard Scaler):

To ensure that numerical values (e.g., Age) are on uniform scale, Standard Scaler was employed. This technique centers the values around a mean of 0 with unity variance, and it can be mathematically deliberated using Eq. (4),

$Z=\frac{x-\mu}{\sigma}$              (4)

where, $x$ is the original feature value, $\mu$ is the mean of the feature and $\sigma$ is the standard deviation. This scaling is especially crucial for models like neural networks, where unscaled features can negatively impact convergence during training. Final Dataset Representation: Let $X$ represent the input feature matrix after preprocessing and $Y$ the encoded target vector using the following Eq. (5),

$X=\left\{x_1, x_2, \ldots x_n\right\}, Y=\left\{y_1, y_2, \ldots y_n\right\}, y_i \in\{0,1\}$               (5)

The preprocessing workflow shown in Figure 2 is a systematic way to prepare unrefined raw lung cancer data for modeling with machine learning [18-22]. The workflow begins with data cleaning, where the missing values are dealt with by mean imputation (a method designed to achieve numerical completeness). After the dataset is cleaned, label encoding takes place by using a manual mapping to convert the two binary values for target classes (I.e. NSCLC and SCLC) to3t8arget classes. Next, categorical variables such as gender or smoking status undergo One-Hot Encoding, creating a binary vector representation for each one. Finally, the remaining numerical attributes (I.e. age) undergo Z-score normalization by standard scaling, standardizing, so that each numerical attribute has zero mean and unit variance.

Figure 2. Workflow of pre-processing Z-score normalization

Besides the standard clinical features, a number of derived characteristics were calculated in the preprocessing process to provide a higher diagnostic presentation of the dataset, including Air Pollution Exposure, Genetic Risk Score, and Symptom Score. These properties were derived or combined on the variables of the environmental exposure, hereditary factors, and indicators of symptoms, respectively. All the new numeric features received new numbers that were scaled using the Standard Scaler after the derivation to ensure that the different numbers had equal scales in the entire dataset. This helped in ensuring that the final pre-processed data that will be used to train the model was consistent, complete, and corresponded with the feature list that was given in Table 3. The result of the preprocessing workflow is then a pre-processed dataset, clearly defined and ready for training classification models.

Table 3. Experimental design on dataset of NSCLC and SCLC

3.3 NSCLC classification

Non-Small Cell Lung Cancer (NSCLC) accounts for nearly 85% of lung cancer cases and includes subtypes such as adenocarcinoma (ADC), squamous cell carcinoma (SCC), and large cell carcinoma (LCC). Accurate identification of NSCLC is essential for treatment planning and prognosis. Traditional diagnostic methods such as histopathological or imaging-based manual assessments are often time-consuming and subjective. To overcome these challenges, a hybrid deep learning model combining CNN and BiLSTM is proposed for NSCLC classification using structured clinical and radiomic descriptors. The CNN component efficiently captures local feature interactions among input attributes, while the BiLSTM learns sequential and contextual dependencies across features, enhancing classification accuracy.

Step 1: Convolutional Layer – Local Feature Extraction

The CNN layer acts as a local filter that identifies inter-feature relationships and high-level attribute combinations within the input feature vector. Each convolutional operation extracts localized correlations (e.g., between age, smoking index, lesion density, and other radiomic scores) using learnable kernels, as defined in Eq. (6):

$f_i=\operatorname{RELU}\left(K * x_i+b\right)$               (6)

where, $f_i$ is the Output feature map for the $i$-th sample, $x_i$ is the Pre-processed input vector, Convolutional kernel can be denoted as $K$, Bias term can be denoted as $b$, $*$ is the Convolution operator and $\operatorname{RELU}(z)=\max (0, z)$ is the activation function introducing non-linearity. This operation captures localized attribute dependencies from structured data.

Step 2: Bidirectional LSTM Representation

The BiLSTM layer captures bidirectional relationships among features — for instance, how clinical attributes such as age or smoking history correlate with radiomic parameters across the feature sequence. This improves contextual understanding by processing information in both forward and backward directions, represented in Eq. (7):

$h_t=\operatorname{BILSTM}\left(f_i\right)=h \underset{t}{\stackrel{\rightarrow}{\rightarrow}} ⊕h \underset{t}{\stackrel{\leftarrow}{\leftarrow}}$               (7)

where, $f_i$ is the CNN output feature map, $\vec{h} \rightarrow$ and $\overleftarrow{h_t}$ are the forward and backward hidden states, $\oplus$ is the Concatenation and $h_t$ is the contextual representation at timestep $t$.

Step 3: Dense Layer – Class Probability Estimation

The dense layer consolidates the extracted features to produce class probabilities. After capturing inter-feature correlations and sequential dependencies, the dense layer computes the likelihood of each class (NSCLC or SCLC) using the sigmoid function in Eq. (8):

$\hat{v}_i=\sigma\left(W h_t+b\right)$             (8)

where, $h_t$ is the Context vector from BiLSTM, $W$ is the weight matrix, $(z)=1 / 1+e-z$ is the Sigmoid function, and $\hat{y}_i\in(0,1)$ is the predicted probability of class SCLC (1). If a threshold (e.g., 0.5) is applied, if $\hat{y}_i<0.5$, classify as NSCLC, else: classify as SCLC.

Step 4: Loss Function – Binary Cross-Entropy

The model is trained using binary cross-entropy, measuring the divergence between predicted and true labels and the binary Cross-Entropy can be mathematically deliberated using Eq. (9)

$L=\left[y_i \log \left(\hat{y}_i\right)+\left(1-y_i\right) \log \left(1-\hat{y}_i\right)\right]$             (9)

where, $y_i$ is the True label $(0=\mathrm{NSCLC}, 1=\mathrm{SCLC})$, $\hat{y}_i$ is the predicted probability, and $L$ is the minimized loss during training.

Figure 3. CNN–BiLSTM architecture for lung cancer classification

The fusion architecture enables the model to learn deeper patterns beyond simple feature associations and is particularly effective in identifying subtle traits of NSCLC that may not be captured by traditional models or shallow classifiers. The CNN–BiLSTM hybrid model for lung cancer diagnosis begins with pre-processed input features, including clinical attributes such as age, gender, and smoking status, along with radiomic descriptors derived from CT analysis. These structured input features are first passed into the CNN block, which identifies localized feature interactions and non-linear relationships among the clinical and radiomic variables associated with NSCLC. The extracted feature representations are then fed into the BiLSTM layer, which models bidirectional dependencies between features, allowing the network to understand complex inter-feature patterns — for instance, how patient age and smoking index jointly influence certain radiomic characteristics or disease tendencies. The resulting contextual representations are then processed through a dense layer to estimate class probabilities. Finally, a sigmoid activation function produces a score between 0 and 1: if the score is less than 0.5, the sample is classified as NSCLC (class 0); otherwise, it is classified as SCLC (class 1). This classification outcome enables the model to distinguish NSCLC cases effectively based on multi-dimensional correlations within clinical and radiomic features, rather than relying on explicit pixel- or image-based patterns. Architecture for the CNN–BiLSTM Lung Classifier is shown in Figure 3.

3.4 SCLC classification in lung cancer diagnosis

This architecture combines Dense Neural Layers for hierarchical abstraction1and a Gated Recurrent Unit (GRU) with an Attention Mechanism, enhancing awareness of the most informative and differentiated patterns from this structured and sequential clinical dataset to help differentiate subtle cases of SCLC. The Att-DGRU model is designed primarily for the identification of likely cases of SCLC that are typically more aggressive and harder to identify. SCLC has a symptom set that often overlaps with other lung cancers, usually leading to a longer diagnostic determination if relying solely on the case history of a patient. The dense layer of the model represents the initial transformation of clinical features both derived from the prospective clinical features as well as those derived from imaging. GRUs are less computationally demanding than LSTMs, so the model can preserve sequential dependencies found in patient record or feature data. The attention mechanism highlights significant or unique features, where features such as rapid progression, smoker status, or certain radiomic signatures can be prioritized, making the model more user-friendly and resilient against over-parameterization.

Let $X=\{x_1, x_2, \ldots$,} represent the pre-processed input feature matrix from clinical and radiomic attributes.

Step 1: Dense Layer for Feature Abstraction

A dense (fully connected) layer is used for initial non-linear transformation of features and it can be mathematically expressed using Eq. (10),

$z_i=\operatorname{ReLU}\left(W_1 x_i+b_1\right)$             (10)

where, $x_i$ be the input feature vector, $W_1$ and $b_1$ are the Learnable weights and bias. Moreover, $\operatorname{Re}(z)=\max (0, z)$.

Step 2: GRU for Sequential Pattern Modeling

The GRU captures time-dependent or ordered feature relations with fewer parameters than LSTM and it can be mathematically expressed using Eq. (11),

$h_t=\operatorname{GRU}\left(z_i, h_{t-1}\right)$             (11)

where, $h_{t-1}, z t$ is the input to GRU and GRU uses update gate and reset gate internally to control flow of information.

Step 3: Attention Mechanism for Informative Focus

Attention allows the model to weigh critical time steps or features more heavily and the Attention Mechanism can be mathematically given in Eqs. (12)-(13),

$t=\sum_{k=1}^T \exp \left(e_t\right) / \exp \left(e_k\right)$             (12)

$C=\sum_{t=1}^T \alpha t h t$             (13)

where, $e_t=v t \tanh (W_2 h t+b_2), \alpha t$ is the attention weight at time $t, e_t$ is the importance score and the Context vector (weighted sum of hidden states) can be denoted as $c$.

Step 4: Output Layer – Sigmoid for Binary Classification

The output layer of the deep learning model employs a sigmoid activation function to map the learned feature representation to a probability score between 0 and 1, indicating the likelihood that a case is classified as Small Cell Lung Cancer (SCLC). This layer accepts a context vector—created through attention and GRU encoding— which is passed through a dense transformation and then the sigmoid activation function which serves to biases all input mapped within the [0,1] interval. If the predicted score is evaluated as ≥ 0.5, the model assigns a case to the SCLC class, if the score is less than 0.5 then the case is assigned to the NSCLC class. This kind of evaluation ensures to provide an unambiguously interpretable binary decision boundary that is clinically relevant. Furthermore, Binary Classification has been given in Eq. (14),

$\hat{y}=\sigma\left(W_3 C+b_3\right)$             (14)

where, $\hat{y} \in(0,1)$ is the Predicted probability of SCLC, $W_3$ and $b_3$ are the Weights and bias for final layer as well as the Sigmoid activation function can be denoted as $\sigma$. This model outputs a probability score where $\hat{y} \geq 0.5$ is classified as SCLC, and $\hat{y}<0.5$ as NSCLC. By integrating attention, it improves the sensitivity of the model towards subtle, discriminative cues typical of SCLC, helping clinicians in early, accurate detection.

Figure 4. Att-DGRU-based lung cancer classification model

Figure 4 outlines the architecture of the Att-DGRU model for lung cancer classification. In this case, the model takes the pre-processed input data, which was passed through a dense layer to project the features into a space for sequential learning. The output was passed to the Gated Recurrent Unit (GRU layer) to capture the short-term dependencies, then the Bidirectional GRU (BiGRU) to improve feature representation, as it allows the model to observe the input sequence in both forward and backward ways. The output from the GRU layers is then passed on to an attention mechanism, which refines it further by emphasizing only the most relevant features for dynamic feature weighting to improve interpretability and classification. The attention-weighted representation is then passed through a fully connected layer to learn the decision boundaries. Subsequently, a softmax layer is used to output a classified outcome that will show probabilities for each lung cancer subtype (e.g., NSCLC or SCLC). This allows for accurate and interpretable diagnosis.