AxonBuild - AI & Full Stack Solutions

Introduction

We developed a comprehensive multilingual NLP solution that transforms unstructured Spanish hospital reviews into actionable business intelligence. By combining fine-tuned transformer models with advanced topic modeling and sentiment analysis, we enabled healthcare administrators to extract granular insights from hundreds of patient reviews, identifying both sentiment patterns and thematic pain points across hospital operations.

Challenge

Healthcare organizations struggle to extract meaningful insights from vast volumes of unstructured patient feedback in Spanish. Traditional manual review processes are time-intensive, subjective, and fail to identify systematic patterns across operational domains like parking, staff interactions, and equipment quality. Without systematic sentiment analysis and thematic categorization, hospital leadership lacks the granular data needed to prioritize operational improvements and allocate resources effectively. The complexity of Spanish language nuances, regional dialects, and healthcare-specific terminology further compounds the challenge, requiring specialized NLP approaches that generic sentiment tools cannot address.

Solution

We created an integrated healthcare intelligence platform that delivers both emotional context and operational insights simultaneously.

Interactive Data Visualization

We built comprehensive dashboards using Plotly and Seaborn, featuring sentiment distribution heatmaps, topic-wise sentiment breakdowns, and temporal trend analysis. Interactive visualizations included scatter plots mapping sentiment intensity across topics, word cloud generations for thematic understanding, and correlation matrices showing cross-topic sentiment relationships.

Embeddings and Semantic Analysis

Our pipeline leveraged pre-trained Spanish BERT embeddings (768-dimensional vectors) for semantic similarity computation and contextual understanding. We implemented cosine similarity metrics for review clustering and used dimensionality reduction via UMAP for visualization-ready representations.

Advanced Sentiment Classification

We implemented a RoBERTa-based transformer model specifically fine-tuned for Spanish healthcare sentiment analysis. The model was trained on the client's labeled dataset using transfer learning techniques, achieving domain-specific accuracy in medical context understanding. We employed PyTorch and Transformers library for model architecture, with hyperparameter optimization using Optuna for learning rate scheduling and batch size tuning.

Production-Ready Automation Pipeline

We architected a fully automated Jupyter notebook pipeline with modular Python classes for data ingestion, preprocessing, model inference, and visualization generation. The solution includes automated data validation, error handling, and configurable parameters for different hospital networks.

Intelligent Topic Discovery and Clustering

We deployed advanced topic modeling using Latent Dirichlet Allocation (LDA) combined with BERT-based embeddings for semantic clustering. Spanish-specific BERT models (BETO/RoBERTa-es) generated dense vector representations, enabling accurate thematic grouping across hospital operational domains. K-means clustering with silhouette analysis determined optimal topic numbers, while coherence scoring validated topic quality.

Technical Approach

Spanish NLP:Fine tuned RoBERTa models for healthcare specific sentiment classification

Topic Modeling:LDA combined with BERT embeddings for thematic discovery

Deep Learning:PyTorch framework with Transformers library for model architecture

Visualization:Plotly and Seaborn for interactive dashboard creation

Automation:Jupyter notebook pipeline with modular Python classes

Deployment:Docker containerization for consistent cross environment deployment

Stack:Python backend, PyTorch ML framework, specialized Spanish BERT models

What we've accomplished

Successfully delivered a comprehensive multilingual NLP analytics platform that transforms Spanish healthcare feedback into strategic intelligence. The solution enables healthcare networks to scale insights across multiple facilities while maintaining cultural and linguistic accuracy through specialized Spanish language models.

Results & Impact

87%

Topic Coherence

Hospital Locations

92%

Sentiment Accuracy

22%

Parking Satisfaction

95%

Processing Efficiency

82%

Staff Interaction Satisfaction

Project Narrative

Rather than treating sentiment analysis and topic modeling as separate analytical tasks, we created an integrated healthcare intelligence platform that delivers both emotional context and operational insights simultaneously. The fine-tuned Spanish language models ensured cultural and linguistic accuracy, while the automated pipeline architecture enables healthcare networks to scale insights across multiple facilities seamlessly. This end-to-end solution demonstrates how advanced NLP can revolutionize healthcare operations by converting unstructured patient voices into precise, actionable business intelligence that directly improves patient experience and operational efficiency.