A Hybrid Hierarchical 1D-CNN-BiLSTM Framework for Extractive Summarization of Biomedical and Clinical Text

# A Hybrid Hierarchical 1D-CNN-BiLSTM Framework for Extractive Summarization of Biomedical and Clinical Text
Source: [https://arxiv.org/html/2609.13481](https://arxiv.org/html/2609.13481)
Muhammad SaifAffiliation:CS Department NUCES\-FAST Lahore, Pakistan l230969@lhr\.nu\.edu\.pkAli Hassan KhanAffiliation:CS Department NUCES\-FAST Lahore, Pakistan l230924@lhr\.nu\.edu\.pkManzer AbbasAffiliation:CS Department NUCES\-FAST Lahore, Pakistan l230970@lhr\.nu\.edu\.pkAffiliation:Hajra WaheedAffiliation:CS Department NUCES\-FAST Lahore, Pakistan hajra\.waheed@nu\.edu\.pkAffiliation:

###### Abstract

Large language models have made abstractive summarization remarkably fluent, but fluency has come at a cost\. Models trained to generate freely over long documents often hallucinate, producing summaries that read well but drift from what the source actually says\. This risk is especially serious in domains such as biomedical and clinical text, where a fabricated detail is not a stylistic flaw but a factual failure\. We address this by removing generation from the pipeline entirely rather than trying to constrain it\. Our Hybrid Hierarchical CNN\-LSTM Summarizer reframes summarization as extractive sentence selection: a stacked, multi\-kernel convolutional network first composes sentence\-level embeddings into progressively richer inter\-sentence factual units, layer by layer, and a bidirectional LSTM then models long\-range dependencies and ordering across the full document\. A lightweight scoring head projects this contextual representation into a per\-sentence importance score, trained end\-to\-end with binary cross\-entropy against oracle extractive labels\. At inference time, a dynamic mean\-plus\-standard\-deviation threshold, with a top\-3 fallback, selects sentences directly from the source document, which are then chronologically reordered into the final summary\. Because every output sentence is copied verbatim from the input, factual drift is structurally impossible rather than merely discouraged\. Evaluated on the open\-domain PubMed benchmark, our architecture outperforms isolated CNN and LSTM baselines, with ablations confirming that wider convolutional receptive fields improve sentence scoring\. On clinical datasets \(MIMIC\-CXR, MIMIC\-IV BHC\), the model excels on unstructured narratives, though it mechanically defaults to simple positional baselines on highly templated reports\. These results suggest that structural constraints, rather than larger models or better decoding heuristics, offer a more reliable path toward factually grounded summarization\. We position this work as a step toward summarization systems that are trustworthy by design rather than by correction\.

###### Index Terms:

Extractive Text Summarization, Clinical Natural Language Processing, Hierarchical 1D\-CNN, Bidirectional LSTM, Electronic Health Records \(EHR\), Factuality Preservation, Biomedical Document Extraction\.

## IIntroduction

Text summarization is broadly categorized into two main approaches: extractive and abstractive\. Extractive summarization identifies and selects key sentences directly from the source text, and is safe by construction, since every word in the output already existed in the input, though it tends to produce summaries that read as disconnected fragments rather than coherent prose\[[1](https://arxiv.org/html/2609.13481#bib.bib1)\]\. Abstractive methods, especially those built on modern large language models, instead read the source and generate an entirely new summary, producing fluent, human\-like text that extractive methods cannot match\[[2](https://arxiv.org/html/2609.13481#bib.bib2)\]\.

That fluency comes at a cost\. When attention spreads thin across long inputs, small inconsistencies compound: the model does not fail by refusing to answer, but by answering confidently with summaries that are fluent, plausible, and factually wrong\. This behavior, commonly called hallucination, is especially dangerous in domains such as biomedical and clinical text, where an invented detail is not a stylistic imperfection but a factual error with real consequences\[[2](https://arxiv.org/html/2609.13481#bib.bib2)\]\.

Much of the recent literature manages hallucination after the fact, through better decoding strategies, factuality scoring, or post\-hoc verification\. We take a different position: rather than generating a summary and then checking whether it is faithful, we ask whether free\-form generation is necessary at all\. Prior work has shown that LSTM\-based sentence classifiers built on pretrained embeddings are effective for identifying salient content in long documents\[[3](https://arxiv.org/html/2609.13481#bib.bib4)\], that stacking recurrent and convolutional components is highly complementary in specialized domains such as biomedical transcripts\[[4](https://arxiv.org/html/2609.13481#bib.bib5)\], and that explicitly modeling a document’s local and hierarchical structure improves sentence selection on long, well\-organized text\[[5](https://arxiv.org/html/2609.13481#bib.bib8)\]\. What none of this work does is combine a sequentially stacked, multi\-kernel hierarchical convolutional network with a bidirectional LSTM and a confidence\-calibrated, dynamically thresholded selection mechanism, evaluated jointly across open\-domain scientific and high\-stakes clinical narrative text\.

We address this gap with a Hybrid Hierarchical CNN\-LSTM Summarizer that reframes summarization as pure sentence selection rather than text generation, so that factual drift is structurally impossible rather than merely discouraged\. The primary contributions of this work are as follows:

1. 1\.Hierarchical Architecture:A sequential 1D\-CNN pipeline that progressively refines local sentence features before modeling global document context via a BiLSTM\.
2. 2\.Adaptive Extraction:A dynamic thresholding strategy that adapts to the unique score distribution of each document, eliminating the need for rigid, fixed\-length summaries\.
3. 3\.Comprehensive Validation:Evaluations across biomedical research and raw clinical narratives, backed by targeted ablation studies that validate our network’s depth and hybrid design\.

## IIRelated Work

### II\-ASurveys of Extractive Summarization

Extractive summarization has been studied from many angles, and two recent reviews give a useful sense of where the field currently stands\. Azam et al\.\[[1](https://arxiv.org/html/2609.13481#bib.bib1)\]surveyed 145 papers and proposed a layered generic architecture for extractive systems, moving through preprocessing, feature extraction, sentence scoring, base modeling, sentence selection, and post\-processing\. Their review found that deep learning methods, including CNN and LSTM hybrids, now dominate across news, scientific, and social media domains\. Zayed et al\.\[[2](https://arxiv.org/html/2609.13481#bib.bib2)\]took a broader view, comparing statistical, graph\-based, and neural approaches across both extractive and abstractive summarization, specifically flagging hallucination, prompt sensitivity, and the lack of standardized cross\-domain benchmarks as open problems\. This directly motivates our work, since limiting hallucination through structural constraints on generation, rather than correcting it after the fact, is the central rationale for our purely extractive approach\.

### II\-BRecurrent and Hybrid CNN\-LSTM Architectures

At the model level, pairing pretrained embeddings with recurrent networks has proven highly effective for extractive tasks\. Framing extraction as binary sentence classification, a simple LSTM paired with BERT embeddings was shown to outperform baseline classifiers on the Cornell Newsroom dataset\[[3](https://arxiv.org/html/2609.13481#bib.bib4)\]\. Comparative studies reinforce this; LSTMs and GRUs have been found to perform at a similarly high level for scientific journal summarization, decisively beating classical latent semantic analysis on ROUGE scores\[[6](https://arxiv.org/html/2609.13481#bib.bib10)\]\.

Building on these recurrent foundations, recent work has explored combining LSTMs with CNNs\. Bedi et al\.\[[4](https://arxiv.org/html/2609.13481#bib.bib5)\]proposed a Deep Dense LSTM network followed by a CNN layer for biomedical transcript summarization, demonstrating strong transferability to medical text\. Similarly, while Transformer\-based models have been found to often edge out standard deep learning baselines on PubMed clinical text\[[7](https://arxiv.org/html/2609.13481#bib.bib6)\], our own results \(Table[I](https://arxiv.org/html/2609.13481#S5.T1)\) show a BERTSUM baseline trailing both variants of our stacked architecture, suggesting the integration of convolutional and recurrent layers remains a robust, lightweight alternative for domain\-specific processing\.

### II\-CHierarchical Structure and Pipeline Complexity

Handling long, structurally complex documents has pushed the field beyond single flat encoders\. Explicitly modeling both local topics and overarching document hierarchy has been shown to outperform flat baselines like BERTSUM on lengthy PubMed and ArXiv papers\[[5](https://arxiv.org/html/2609.13481#bib.bib8)\]\. In multi\-document settings, classical feature scoring has been combined with a CNN\-RNN module to reduce redundancy\[[8](https://arxiv.org/html/2609.13481#bib.bib9)\]\.

Furthermore, as summarization pipelines expand into multimedia and specialized domains, managing structural complexity becomes critical\. A cascaded audio\-visual pipeline for multimedia summarization has been proposed\[[9](https://arxiv.org/html/2609.13481#bib.bib7)\], weak supervision has been combined with a BART\-CNN model for clinical question answering\[[10](https://arxiv.org/html/2609.13481#bib.bib12)\], and extraction has been extended into low\-resource languages such as Urdu\[[11](https://arxiv.org/html/2609.13481#bib.bib11)\]\. However, cascaded and highly modular pipelines risk error propagation, where an early mistake carries forward into the final summary\. This underscores the need for architectures that are shallow, tightly coupled, and supervised end\-to\-end\.

This literature traces a fairly clear path from statistical methods toward deep learning and hybrid CNN\-LSTM architectures\. What is largely missing, however, is a stacked, sequentially hierarchical convolutional stage where each kernel width is applied directly to the previous layer’s feature map rather than in parallel, paired with a bidirectional LSTM and a dynamically thresholded, confidence\-calibrated sentence selector\. Our work builds on the CNN\-LSTM foundation but pushes it toward a lightweight, structurally sound design\. By evaluating jointly across open\-domain scientific text and high\-stakes clinical narratives, we present a system where hallucination is constrained by construction, as the model is never permitted to generate text that did not already exist in the source\.

## IIIMethodology

This research proposes a Hybrid Hierarchical Stacked 1D\-CNN–BiLSTM Extractive Summarizer\. In contrast to conventional multi\-kernel convolutional architectures, which apply several kernel widths in parallel and concatenate their outputs, our design adopts a strictly sequential, stacked hierarchy: each convolutional layer operates directly on the feature map produced by the layer before it, allowing lower\-ordernn\-gram representations to be progressively refined into higher\-order factual units before contextualization and scoring\. The complete pipeline, including tensor dimensions at each stage, is illustrated in Fig\.[1](https://arxiv.org/html/2609.13481#S3.F1)\.

![Refer to caption](https://arxiv.org/html/2609.13481v1/b.drawio.png)Fig\. 1:Overall architecture of the proposed Hybrid Hierarchical Stacked 1D\-CNN\-BiLSTM Summarizer, detailing tensor transformations and sequence dimensions at each pipeline stage\.### III\-ASentence Tokenization and Dense Vector Representation

A documentDDis represented as an ordered sequence ofNNsentences,D=\{s1,…,sN\}D=\\\{s\_\{1\},\\dots,s\_\{N\}\\\}, via NLTK tokenization\. Each sentencesis\_\{i\}is encoded using a domain\-specific Transformer \(SciBERT\), taking its\[CLS\]token as a dense embedding:

𝐞i=TransformerEncoder​\(si\)\[CLS\]∈ℝDembed,Dembed=768\.\\mathbf\{e\}\_\{i\}=\\text\{TransformerEncoder\}\(s\_\{i\}\)\_\{\[\\text\{CLS\}\]\}\\in\\mathbb\{R\}^\{D\_\{\\text\{embed\}\}\},\\quad D\_\{\\text\{embed\}\}=768\.\(1\)Documents are standardized to a fixed maximum length ofNseq=512N\_\{\\text\{seq\}\}=512sentences via zero\-padding or truncation, yielding a document\-level embedding tensor𝐗\(0\)∈ℝB×Nseq×Dembed\\mathbf\{X\}^\{\(0\)\}\\in\\mathbb\{R\}^\{B\\times N\_\{\\text\{seq\}\}\\times D\_\{\\text\{embed\}\}\}, whereBBis the batch size, transposed to a channel\-first layout𝐗\(0\)∈ℝB×768×512\\mathbf\{X\}^\{\(0\)\}\\in\\mathbb\{R\}^\{B\\times 768\\times 512\}prior to convolution\.

### III\-BStacked Hierarchical 1D Convolutional Feature Refinement

The central architectural contribution is a sequentially stacked 1D convolutional network where each kernel operates on the preceding layer’s output \(Fig\.[2](https://arxiv.org/html/2609.13481#S3.F2)\)\. This design encourages progressive compositional abstraction rather than aggregating parallel fixed\-width views of the same input\.

![Refer to caption](https://arxiv.org/html/2609.13481v1/architecture.png)Fig\. 2:Conceptual flow of the hierarchical 1D\-CNN stack\. Lower\-order sentence\-level features are progressively composed into higher\-order factual units layer by layer\.Formally, each layerl∈\{1,2,3,4\}l\\in\\\{1,2,3,4\\\}applies a 1D convolution with kernel widthkl=lk\_\{l\}=ldirectly to the output of the preceding layer:

𝐗\(l\)=ReLU\(Conv1Dk=l\(𝐗\(l−1\)\)\)∈ℝB×128×512,l=1,…,4\.\\mathbf\{X\}^\{\(l\)\}=\\text\{ReLU\}\\left\(\\text\{Conv1D\}\_\{k=l\}\\left\(\\mathbf\{X\}^\{\(l\-1\)\}\\right\)\\right\)\\in\\mathbb\{R\}^\{B\\times 128\\times 512\},\\quad l=1,\\dots,4\.

\(2\)
All layers useCfilters=128C\_\{\\text\{filters\}\}=128filters withpadding=’same’to preserve the sequence length throughout the stack\. Because convolution is applied over a sequence of sentence\-level embeddings, kernel widths represent local inter\-sentence context rather than conventional word\-level n\-grams\. Layer 1 \(k=1k=1\) projects the 768\-dimensional sentence embeddings into a 128\-dimensional single\-sentence feature space; Layer 2 \(k=2k=2\) composes adjacent sentence features into a 2\-sentence contextual window; Layer 3 \(k=3k=3\) further composes this into 3\-sentence\-level coarse factual units; and Layer 4 \(k=4k=4\) aggregates these into broader, document\-level factual units, completing the hierarchical composition\. Because each layerlloperates on the representation learned at layerl−1l\-1rather than on the raw input independently, this stacked formulation is the key structural distinction between our design and standard parallel multi\-kernel CNNs\.

### III\-CInter\-Sentence Contextual Modeling via BiLSTM

While the stacked convolutional stack captures increasingly abstract local factual structure, extractive summarization additionally requires modeling long\-range dependencies and discourse\-level ordering across the document\. The final feature map𝐗\(4\)\\mathbf\{X\}^\{\(4\)\}is transposed to a sequence\-first layout,𝐇cnn=\(𝐗\(4\)\)T∈ℝB×512×128\\mathbf\{H\}\_\{\\text\{cnn\}\}=\(\\mathbf\{X\}^\{\(4\)\}\)^\{T\}\\in\\mathbb\{R\}^\{B\\times 512\\times 128\}, and passed through a single\-layer Bidirectional LSTM with hidden dimensionhdim=128h\_\{\\text\{dim\}\}=128per direction\. The forward and backward hidden states are concatenated at each sentence position to produce a256256\-dimensional contextualized representation,𝐇lstm∈ℝB×512×256\\mathbf\{H\}\_\{\\text\{lstm\}\}\\in\\mathbb\{R\}^\{B\\times 512\\times 256\}, which jointly encodes each sentence’s local factual content and its position within the surrounding document flow\.

### III\-DExtractive Scoring Head and Training Objective

Each contextualized sentence representation is mapped to a scalar saliency score via a linear projection followed by a sigmoid activation,y^i=σ⁡\(𝐖c​𝐇lstm,i\+bc\)∈\[0,1\]\\hat\{y\}\_\{i\}=\\sigma\(\\mathbf\{W\}\_\{c\}\\mathbf\{H\}\_\{\\text\{lstm\},i\}\+b\_\{c\}\)\\in\[0,1\], wherey^i\\hat\{y\}\_\{i\}denotes the predicted probability that sentencesis\_\{i\}belongs in the summary\. The model is trained end\-to\-end using masked binary cross\-entropy over valid sentences against oracle labelsyi∈\{0,1\}y\_\{i\}\\in\\\{0,1\\\}; padded positions are excluded from the loss via a binary mask\.

### III\-EDynamic Confidence\-Calibrated Inference

Rather than a fixed top\-kkcutoff applied uniformly across documents of varying length and content density, we select sentences at inference time using a per\-document, confidence\-calibrated threshold:

τ=μ⁡\(𝐲^valid\)\+0\.5​σ​\(𝐲^valid\),𝒮selected=\{si\|y^i\>τ\}\.\\begin\{gathered\}\\tau=\\mu\(\\hat\{\\mathbf\{y\}\}\_\{\\text\{valid\}\}\)\+0\.5\\,\\sigma\(\\hat\{\\mathbf\{y\}\}\_\{\\text\{valid\}\}\),\\\\ \\mathcal\{S\}\_\{\\text\{selected\}\}=\\left\\\{s\_\{i\}\\;\\middle\|\\;\\hat\{y\}\_\{i\}\>\\tau\\right\\\}\.\\end\{gathered\}\(3\)
whereμ\\muandσ\\sigmaare the mean and standard deviation of the predicted scores over theNvalidN\_\{\\text\{valid\}\}non\-padded sentences in that document\. In degenerate cases where fewer than two sentences exceedτ\\tau, a fallback rule selects the topk=min⁡\(3,Nvalid\)k=\\min\(3,N\_\{\\text\{valid\}\}\)highest\-scoring sentences\. Selected sentences are finally reordered according to their original position in the source document, preserving narrative and chronological coherence in the extracted summary\.

## IVExperimental Setup

In this section, we describe the datasets, baseline models, training protocol, and evaluation metrics used to empirically validate our proposed architecture\.

### IV\-ADatasets and Benchmark Suite

We evaluate our model across three diverse biomedical and clinical datasets to test both open\-domain scientific literature summarization and high\-stakes clinical EHR narratives:

- •PubMed:A large\-scale biomedical document summarization dataset\[[12](https://arxiv.org/html/2609.13481#bib.bib13)\]\. The model was trained on roughly 10,000 documents\. For computational feasibility, evaluation used a fixed subset of 300 test documents, with ground\-truth abstracts serving as target summaries\.
- •MIMIC\-CXR:A specialized clinical dataset containing chest X\-ray radiology reports\[[13](https://arxiv.org/html/2609.13481#bib.bib14)\]\. Summarization here requires extracting critical diagnostic findings and impressions from semi\-structured radiological notes\.
- •MIMIC\-IV BHC \(Brief Hospital Course\):A complex clinical narrative dataset derived from electronic health records\[[14](https://arxiv.org/html/2609.13481#bib.bib15)\]\.

### IV\-BBaseline Models and Architectural Variants

To isolate the individual contributions of our stacked convolutional layers, recurrent modeling, pre\-trained backbones, and overall pipeline, we compare our full architecture against several standard baselines and ablation variants:

- •Lead\-3 Heuristic:A widely used non\-trainable baseline that blindly extracts the first three sentences of a document\.
- •TextRank:A graph\-based, unsupervised extractive baseline that constructs sentence graphs using TF\-IDF cosine similarity and ranks sentences via PageRank\.
- •BERTSUM:A pretrained BERT\-based extractive summarization baseline that scores and selects sentences via a fine\-tuned transformer encoder\[[15](https://arxiv.org/html/2609.13481#bib.bib3)\]\.
- •CNN\-Only Model:A variant that passes sentence embeddings through the 4\-layer stacked CNN stack followed directly by a linear classification head, removing the BiLSTM sequential bottleneck\.
- •LSTM\-Only Model:A variant that feeds sentence embeddings directly into the BiLSTM without convolutional feature refinement\.
- •Vanilla BERT Variant:Our complete stacked CNN\-BiLSTM architecture using standardbert\-base\-uncasedembeddings \(Dembed=768D\_\{\\text\{embed\}\}=768\) instead of domain\-specific SciBERT\.
- •Proposed Hierarchical Model \(SciBERT\):Our full framework combining SciBERT embeddings \(allenai/scibert\_scivocab\_uncased\), 4\-layer stacked 1D\-CNNs, BiLSTM, and dynamic confidence thresholding\.

### IV\-CImplementation and Training Details

All models were implemented in PyTorch and trained on Apple Silicon MPS / CUDA GPU hardware\. Input documents were truncated or padded to a maximum sequence length ofNseq=512N\_\{\\text\{seq\}\}=512sentences, with individual sentence representations embedded into a 768\-dimensional hidden space\.

The stacked 1D\-CNN utilizedCfilters=128C\_\{\\text\{filters\}\}=128channels across kernel sizesk∈\{1,2,3,4\}k\\in\\\{1,2,3,4\\\}withpadding=’same’\. The BiLSTM layer operated with a hidden dimension of 128 per direction \(256 concatenated\)\. Models were optimized using Adam with Binary Cross\-Entropy loss \(ℒBCE\\mathcal\{L\}\_\{\\text\{BCE\}\}\) over valid non\-padded sentences\. Performance was evaluated using standardROUGE\-1,ROUGE\-2, andROUGE\-LF1\-scores against ground\-truth summaries using NLTK stemming\. For supervised training, our framework directly utilized the binary extractive labels provided within the benchmark datasets to serve as the oracle\. These binary indicators \(yi∈\{0,1\}y\_\{i\}\\in\\\{0,1\\\}\) explicitly mark which source sentences optimally align with the ground\-truth summary\.

## VResults and Discussion

### V\-AOpen\-Domain Biomedical Benchmark \(PubMed\)

We first evaluate our framework on the open\-domain PubMed benchmark, with the comparative results detailed in Table[I](https://arxiv.org/html/2609.13481#S5.T1)\. Our complete architecture combining the SciBERT embedding space, the hierarchical stacked CNN, and the BiLSTM achieves the highest overall performance across all metrics, reaching a ROUGE\-1 score of0\.7143, a ROUGE\-2 of0\.7090, and a ROUGE\-L of0\.7191\.

TABLE I:Performance Comparison on the PubMed Benchmark DatasetBeyond the absolute scores, the empirical findings on PubMed reveal two critical insights about our architectural design:

1. 1\.The Limitations of Surface\-Level Heuristics:Traditional, unsupervised baselines struggle to navigate the density of lengthy biomedical texts\. Lead\-3 and TextRank max out at ROUGE\-1 scores of 0\.5439 and 0\.5628, respectively\. Our proposed model outperforms Lead\-3 by an absolute ROUGE\-1 gain of\+0\.1704\(a relative improvement of31\.3%\) and TextRank by an absolute gain of\+0\.1515\(a relative improvement of26\.9%\)\.
2. 2\.The Value of Domain\-Aware Embeddings:Swapping a general\-purpose Vanilla BERT encoder for the medically\-tuned SciBERT yields a direct absolute gain of\+0\.0363\(a5\.4%relative improvement\) in ROUGE\-1\. This validates that injecting domain\-specific vocabulary at the very beginning of the pipeline provides a much richer foundational representation for the convolutional layers to build upon\.

### V\-BClinical Narrative Performance \(MIMIC\-CXR & MIMIC\-IV BHC\)

Moving from polished scientific papers to raw electronic health records \(EHRs\) introduces a significant domain shift\. Table[II](https://arxiv.org/html/2609.13481#S5.T2)details our framework’s performance across clinical narratives\. It is important to note that unlike PubMed articles, where target abstracts are often near\-verbatim extracts of the paper, clinical ground\-truth summaries are highly abstractive and human\-synthesized\. This inherent structural difference naturally lowers the absolute ROUGE scores across the board for all extractive models\.

TABLE II:Comparative Benchmarking on MIMIC\-CXR and MIMIC\-IV BHCOn MIMIC\-CXR, the proposed model performs nearly identically to the Lead\-3 baseline, with scores within 0\.002 across all three metrics\. This near\-parity is not coincidental: it reflects a mechanical behavior\. Due to the extreme brevity and structured nature of these radiology reports, very few sentences exceed the dynamic confidence thresholdτ\\tau\. Consequently, the model defaults to its top\-3 fallback rule for nearly every document\. This reveals a critical limitation of our dynamic selection strategy on highly templated, short clinical documents: the fallback mechanism forces the architecture to functionally collapse into a simple positional baseline\. TextRank achieves a marginally higher ROUGE\-1 \(0\.1829\) on this dataset by exploiting high\-frequency clinical keywords, though this form of surface\-level lexical matching is not guaranteed to generalize to less formulaic text\.

This distinction becomes clear on the MIMIC\-IV Brief Hospital Course \(BHC\) dataset, where the predictable radiology template is absent and narratives are longer and less rigidly structured\. Under these conditions, TextRank’s keyword\-driven heuristic degrades sharply, dropping to a ROUGE\-1 of 0\.1353\. In contrast, the proposed model achieves its largest relative advantage on MIMIC\-IV BHC, reaching a ROUGE\-1 of0\.2682\. This translates to an absolute improvement of\+0\.0620\(a30\.1%relative gain\) over Lead\-3 and\+0\.1329\(a98\.2%relative gain\) over TextRank\. Taken together, these results suggest a consistent pattern rather than uniform superiority: on structured clinical text with predictable information placement, a simple positional baseline is already competitive, whereas on unstructured, narrative clinical text, the hybrid hierarchical model’s dynamic thresholding and multi\-level feature refinement provide a clear, measurable advantage over both positional and graph\-based baselines\.

### V\-CAblation Studies: Component Isolation & CNN Depth Progression

To truly understandwhythe hierarchical architecture works, we conducted component and depth ablations on the PubMed dataset to examine the contribution of the hybrid architecture\. We systematically isolated key components and disabled upper convolutional layers to observe how depth impacts feature learning\. The results of component isolation and depth progression are documented in Table[III](https://arxiv.org/html/2609.13481#S5.T3)and Table[IV](https://arxiv.org/html/2609.13481#S5.T4), respectively\.

TABLE III:Ablation Study: Component Isolation and Hybrid Coupling \(PubMed\)#### V\-C1The Synergy of Hybrid Architectures

When evaluated in isolation \(as shown in Table[III](https://arxiv.org/html/2609.13481#S5.T3)\), the CNN\-Only model captures local patterns reasonably well \(ROUGE\-1: 0\.6339\), while the LSTM\-Only model manages sequential flow more effectively \(ROUGE\-1: 0\.6766\)\. However, coupling them into a unified, hierarchical pipeline yields a substantial performance leap to0\.7143\. This represents an absolute gain of\+0\.0377\(5\.6%relative improvement\) over the LSTM\-Only baseline and\+0\.0804\(12\.7%relative improvement\) over the CNN\-Only baseline\. This confirms our core hypothesis: local inter\-sentence factual refinement and global context modeling are strictly complementary forces\.

TABLE IV:Ablation Study: Stacked CNN Depth Progression \(PubMed\)
#### V\-C2The Impact of Stacked Convolutional Depth

The layer progression results in Table[IV](https://arxiv.org/html/2609.13481#S5.T4)tell a compelling story about representation learning\. Relying solely on 1\-sentence or 2\-sentence stacks yields near\-stagnant scores \(ROUGE\-1: 0\.6262 and 0\.6291, respectively\), indicating that medical concepts are too complex to be captured in one or two adjacent statements\. However, the moment we introduce Layer 3 \(3\-sentence window\), the model experiences a sharp absolute gain of\+0\.0489\(a7\.8%relative jump\) in performance\. This suggests that 3\-sentence inter\-relationships provide the critical receptive field required to recognize a fully\-formed medical claim\. Finally, capping the architecture with Layer 4 acts as a global synthesizer, providing a final absolute boost of\+0\.0363\(a5\.4%relative improvement\) and securing the peak score\. This progression firmly validates our design choice: deep, sequential feature refinement is essential for composing robust, high\-level factual units before passing them to a recurrent classifier\.

### V\-DLimitations

While the proposed architecture demonstrates strong empirical performance, our reliance on standard ROUGE metrics may not fully capture the clinical usefulness or deeper factual consistency of the extracted summaries\. Furthermore, the strictly extractive framework can produce fragmented narratives when diagnostic information is dispersed, and its dynamic thresholding strategy remains highly sensitive to underlying document structures\.

## VIConclusion

In this paper, we addressed the critical issue of factual hallucination in biomedical and clinical text summarization by reframing the task as a strictly extractive, confidence\-calibrated sentence selection problem\. We proposed a Hybrid Hierarchical Stacked 1D\-CNN\-BiLSTM framework that structurally prevents generative factual drift while maintaining high contextual fidelity\. By sequentially stacking convolutional layers, the architecture progressively refines dense sentence embeddings into high\-level multi\-sentence factual units, which are subsequently contextualized across the entire document sequence using a Bidirectional LSTM\.

On the open\-domain PubMed benchmark, our complete 4\-layer architecture achieved a peak ROUGE\-1 score of 0\.7143, significantly outperforming non\-trainable heuristics and isolated baseline variants\. Across the two clinical datasets, the model’s behavior was more nuanced: on the highly templated MIMIC\-CXR radiology reports, it performed nearly identically to the Lead\-3 baseline, suggesting that on rigidly structured text the model converges toward a similar positional strategy rather than offering additional benefit; on the less predictable, unstructured MIMIC\-IV BHC narratives, it achieved its largest relative gains over both Lead\-3 and TextRank, indicating that the hierarchical architecture is most useful precisely where simple positional or keyword\-based heuristics break down\.

Ultimately, this work suggests that structural constraints and supervised extractive architectures offer a reliable, computationally efficient, and trustworthy alternative to large autoregressive language models in high\-stakes medical domains, though their advantage over simpler baselines is dataset\-dependent rather than universal\. Future work will explore expanding this hierarchical framework to multi\-document patient history summarization and evaluating cross\-lingual zero\-shot extraction for under\-resourced healthcare settings\.

## References

- \[1\]M\. Azam, S\. Khalid, S\. Almutairi, H\. A\. Khattak, A\. Namoun, A\. Ali, and H\. S\. M\. Bilal\(2025\)Current trends and advances in extractive text summarization: a comprehensive review\.IEEE Access13,pp\. 28150–28166\.Cited by:[§I](https://arxiv.org/html/2609.13481#S1.p1.1),[§II\-A](https://arxiv.org/html/2609.13481#S2.SS1.p1.1)\.
- \[2\]S\. Zayed, M\. Ezzat, and H\. A\. Hefny\(2025\)Automatic text summarization: a review of approaches, challenges, and future directions\.Journal of Computer Science & Technology25\.Cited by:[§I](https://arxiv.org/html/2609.13481#S1.p1.1),[§I](https://arxiv.org/html/2609.13481#S1.p2.1),[§II\-A](https://arxiv.org/html/2609.13481#S2.SS1.p1.1)\.
- \[3\]S\. Biswas, M\. Biswas, A\. Mandal, F\. T\. Liza, and J\. Sarker\(2025\)Efficient extractive text summarization for online news articles using machine learning\.arXiv preprint arXiv:2509\.15614\.Cited by:[§I](https://arxiv.org/html/2609.13481#S1.p3.1),[§II\-B](https://arxiv.org/html/2609.13481#S2.SS2.p1.1)\.
- \[4\]P\. P\. S\. Bedi, M\. Bala, and K\. Sharma\(2024\)Extractive text summarization for biomedical transcripts using deep dense lstm\-cnn framework\.Expert Systems41\(7\),pp\. e13490\.Cited by:[§I](https://arxiv.org/html/2609.13481#S1.p3.1),[§II\-B](https://arxiv.org/html/2609.13481#S2.SS2.p2.1)\.
- \[5\]T\. Wang, C\. Yang, M\. Zou, J\. Liang, D\. Xiang, W\. Yang, H\. Wang, and J\. Li\(2024\)A study of extractive summarization of long documents incorporating local topic and hierarchical information\.Scientific Reports14\(1\),pp\. 10140\.Cited by:[§I](https://arxiv.org/html/2609.13481#S1.p3.1),[§II\-C](https://arxiv.org/html/2609.13481#S2.SS3.p1.1)\.
- \[6\]D\. Fitrianah and R\. N\. Jauhari\(2022\)Extractive text summarization for scientific journal articles using long short\-term memory and gated recurrent units\.Bulletin of Electrical Engineering and Informatics11\(1\),pp\. 150–157\.Cited by:[§II\-B](https://arxiv.org/html/2609.13481#S2.SS2.p1.1)\.
- \[7\]N\. Dattaet al\.\(2024\)Extractive text summarization of clinical text using deep learning models\.In2024 Second International Conference on Emerging Trends in Information Technology and Engineering \(ICETITE\),pp\. 1–6\.Cited by:[§II\-B](https://arxiv.org/html/2609.13481#S2.SS2.p2.1)\.
- \[8\]B\. Purushothaman and G\. Tamilpavai\(2026\)Enhancing multi\-document summarization with hybrid convolutional recurrent neural networks for efficient information extraction\.Neural Computing and Applications38\(7\),pp\. 205\.Cited by:[§II\-C](https://arxiv.org/html/2609.13481#S2.SS3.p1.1)\.
- \[9\]T\. Hossain, A\. Islam, M\. S\. Hossain, and A\. A\. Rasel\(2025\)A cascaded architecture for extractive summarization of multimedia content via audio\-to\-text alignment\.arXiv preprint arXiv:2504\.06275\.Cited by:[§II\-C](https://arxiv.org/html/2609.13481#S2.SS3.p2.1)\.
- \[10\]N\. Saeed\(2025\)Medifact at peranssumm 2025: leveraging lightweight models for perspective\-specific summarization of clinical q&a forums\.InProceedings of the Second Workshop on Patient\-Oriented Language Processing \(CL4Health\),pp\. 331–339\.Cited by:[§II\-C](https://arxiv.org/html/2609.13481#S2.SS3.p2.1)\.
- \[11\]S\. Nazir, M\. Asif, S\. Ahmad, H\. Aljuaid, and S\. Ahmad\(2026\)Enhanced extractive text summarization framework for low\-resourced urdu language\.PLoS One21\(2\),pp\. e0341596\.Cited by:[§II\-C](https://arxiv.org/html/2609.13481#S2.SS3.p2.1)\.
- \[12\]A\. Cohan, F\. Dernoncourt, S\. Doo, C\. Wu, P\. Bhashyam, J\. Guo, and N\. Goharian\(2018\)A discourse\-aware attention model for abstractive summarization of long documents\.InProceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies \(NAACL\-HLT\),pp\. 615–621\.Cited by:[1st item](https://arxiv.org/html/2609.13481#S4.I1.i1.p1.1)\.
- \[13\]A\. E\. Johnson, T\. J\. Pollard, S\. J\. Berkowitz, N\. R\. Greenbaum, M\. P\. Lungren, C\. Deng, R\. G\. Mark, and S\. Horng\(2019\)MIMIC\-cxr, a de\-identified publicly available database of chest radiographs with radiologist reports\.Scientific Data6\(1\),pp\. 317\.Cited by:[2nd item](https://arxiv.org/html/2609.13481#S4.I1.i2.p1.1)\.
- \[14\]A\. E\. Johnson, L\. Bulgarelli, L\. Shen, A\. Gayles, A\. Shammout, S\. Horng, T\. J\. Pollard, S\. Moodie, B\. Gow, L\. H\. Lehman,et al\.\(2023\)MIMIC\-iv, a freely accessible electronic health record dataset\.Scientific Data10\(1\),pp\. 1\.Cited by:[3rd item](https://arxiv.org/html/2609.13481#S4.I1.i3.p1.1)\.
- \[15\]Y\. Liu\(2019\)Fine\-tune bert for extractive summarization\.InarXiv preprint arXiv:1903\.10318,Cited by:[3rd item](https://arxiv.org/html/2609.13481#S4.I2.i3.p1.1)\.