Introduction
Predicting the right treatment for a patient requires a comprehensive understanding of their entire medical journey. Modeling years of clinical history—where a diagnosis from five years ago might dictate a prescription today—presents a significant computational challenge. Researchers from Deakin University have introduced DCw-MANN (Dual Controller write-protected Memory Augmented Neural Network), a specialized architecture designed to navigate the long-term dependencies within Electronic Medical Records (EMRs).
Methodology
The core innovation of DCw-MANN lies in adapting the Differentiable Neural Computer (DNC) for healthcare by introducing two specialized controllers and a "write-protect" mechanism.
Dual Controllers
Instead of a single neural controller, the model utilizes two LSTMs:
- Encoder ($LSTM_E$): Processes the patient's history and writes relevant information into the external memory.
- Decoder ($LSTM_D$): Generates the treatment sequence by reading from the memory.
Write-Protected Memory
During the decoding phase, the memory is strictly read-only. This prevents the decoder from "polluting" stored medical history with its own intermediate predictions, ensuring recommendations remain grounded in the actual patient record.
Experimental Setup
The model was evaluated on two primary datasets:
- MIMIC-III: A large-scale dataset of de-identified health data from over 40,000 critical care patients. Tasks included procedure prediction and medication prescription.
- Synthetic Odd-Even Task: A custom task requiring the model to remember a sequence of odd numbers to predict corresponding even numbers, testing its ability to handle algorithmic dependencies.
Results
DCw-MANN consistently outperformed baselines, including standard Seq2Seq models and vanilla DNCs.
Medication Prescription (MIMIC-III)
On this challenging task, DCw-MANN achieved a Precision of 0.598 and a Jaccard Coefficient of 0.556. This significantly outperformed the vanilla DNC (0.577 / 0.529) and nearly tripled the performance of standard Seq2Seq models (0.220 / 0.138).
Synthetic Task Performance
The architectural improvements were further validated on the Odd-Even task. DCw-MANN achieved a Normalized Levenshtein Distance (NLD) of 0.082, a substantial improvement over the standard DNC's 0.267.
Discussion and Conclusion
The success of DCw-MANN highlights that specialization in controller functions—separating encoding and decoding—is more effective for complex medical sequences. Furthermore, preventing memory overwrites during prediction preserves historical integrity. While currently focused on code prediction, future work could extend this architecture to multi-task learning (e.g., mortality prediction) and advanced medical question-answering systems.
Summary updated