Automated Disease Prediction and Prescription System with Regional Kannada Synonym Integration

Technical Summary: Automated Disease Prediction and Prescription System with Regional Kannada Synonym Integration

Problem Statement
Effective healthcare delivery relies heavily on clear communication between providers and patients. In multilingual regions, language barriers frequently lead to misdiagnoses, medication errors, and suboptimal patient outcomes. While Neural Machine Translation (NMT) has advanced significantly, existing solutions like Google Translate and Whisper often fail to accurately translate low-resource languages, particularly when dealing with regional dialects and medical synonyms. In Karnataka, India, patients often describe the same disease using different regional Kannada terms (synonyms) that standard translation tools misinterpret or translate literally (e.g., translating a regional term for "itching" as "breakfast" or "irritation"). Furthermore, the lack of public, large-scale medical datasets for these specific language pairs limits the development of robust, domain-adaptable translation models.

Methodology
The authors propose an automated system that integrates speech-to-text conversion, regional synonym-aware machine translation, and disease prediction. The core technical pipeline involves:

1. Data Collection and Pre-processing: A custom dataset of approximately 900 Kannada-English sentence pairs was created, focusing on patient symptoms and disease descriptions. The data underwent cleaning (lowercasing, removing special characters), tokenization, and vocabulary creation.
2. Model Architecture: The system employs a Recurrent Neural Network (RNN) architecture, specifically utilizing Long Short-Term Memory (LSTM) units to handle sequential data and long-term dependencies.
   • Encoder-Decoder Framework: An encoder processes the input Kannada text into a fixed-dimensional vector (hidden and cell states), while a decoder generates the English translation.
   • Attention Mechanisms: The model integrates both general and global attention mechanisms to allow the decoder to dynamically focus on relevant parts of the source sentence, mitigating the information bottleneck of standard encoder-decoder models.
   • Training Strategy: The Teacher Forcing algorithm is used during training to provide the decoder with precise reference outputs, improving convergence.
   • Decoding: Both Greedy Decoding and Beam Search approaches were evaluated to optimize translation fluency and accuracy.
3. Synonym Integration: A key component of the methodology is the system's ability to recognize regional Kannada synonyms for diseases. Unlike standard translators that may fail on dialectal variations, this model is trained to map these regional variations to the correct medical terminology in English.
4. Prescription Prediction: Once the text is translated and the disease is identified, the system queries a dataset to predict the corresponding medication.

Key Contributions

• Regional Synonym Handling: The development of a translation model specifically designed to handle Kannada-English pairs with a focus on regional medical synonyms, addressing a gap where tools like Google Translate frequently fail.
• Architectural Implementation: The integration of attention mechanisms and the teacher forcing algorithm within an RNN-LSTM framework to enhance translation quality for low-resource languages.
• Comparative Evaluation: A comprehensive comparison of the proposed model against Google Translate and Whisper translator, demonstrating superior performance in specific medical contexts.
• Dataset Creation: The creation of a specialized dataset of ~900 patient symptom sentences to facilitate training and evaluation in this domain.

Results
The proposed system was evaluated using standard NMT metrics: BLEU, METEOR, and Translation Edit Rate (TER).

• Translation Accuracy: The proposed model achieved a BLEU score range of 41.5–31.8 and a METEOR score of 52.2%, outperforming Google Translator (BLEU 35.8–28.4, METEOR 39.5%) and Whisper Translator (BLEU 39.8–30.8, METEOR 41.5%).
• Error Reduction: The model demonstrated a lower Translation Edit Rate (TER) of 53.1% compared to 69.2% for Google and 64.6% for Whisper.
• Synonym Recognition: In qualitative tests, the proposed model correctly identified regional terms for diseases such as "Itching" (translating various regional Kannada terms correctly), "Asthma," and "Varicose Veins," whereas Google and Whisper produced literal or incorrect translations (e.g., translating a regional term for itching as "Breakfast").
• Prescription Prediction: The system reported a 95% accuracy in predicting medications based on the translated disease descriptions, though this is assessed against prediction errors within the specific dataset context.

Significance and Claims
The paper claims that the system successfully bridges the communication gap in healthcare for Kannada-speaking populations by accurately translating regional dialects and synonyms that standard tools miss. By enabling robots or automated systems to understand patient inputs in natural language and forecast appropriate medications, the system aims to assist healthcare professionals in managing high patient volumes more efficiently. The authors emphasize that while the system shows promising results, it is designed to support, not replace, medical professionals; all predicted prescriptions are intended to be verified by human specialists to ensure safety, acknowledging the limitations of AI in guaranteeing seamless accuracy.