Adopting Time-Aware Long-Short Term Memory for Psychosis Prognosis Prediction
Summary
Schizophrenia is a complex and heterogeneous disorder because different underlying biological deficits may manifest the same symptoms across individuals. Therefore, the effective treatments can vary from one patient to another. Medicine helpful for one person, may not work or even worsen the condition of another. Individualized and accurate prediction of long-term disease course and therapy response may help navigate treatment decisions. Thus machine learning methods might be useful in treatment outcome prediction. Long-Short Term Memory (LSTM) is a Recurrent Neural Network (RNN) variant capable of handling long-term event dependencies, which are common in medical data. However, it does require regular time intervals between events. In contrast to standard LSTM, Time-Aware LSTM (T-LSTM) can handle and incorporate information about irregular time intervals in data via a time decay function.
The goal of this study was to compare the performance of LSTM and T-LSTM models for the psychosis prognosis prediction (PPP) task, which tries to predict if a patient will be in a remission state. OPTiMiSE dataset, which comes from a clinical study investigating whether switching antipsychotics improves outcomes in first-episode schizophrenia patients, was used in this research. First, we checked if there was any performance difference between LSTM and T-LSTM models. Second, we investigated the effect of adjusting the T-LSTM decay function shape by learning its parameters through the backpropagation procedure. Parametric sigmoid with one and two trainable variables was used as a time decay function.
We managed to improve and obtain more stable results, while adjusting the decay function shape to the PPP application. The area under the curve (AUC) score increased from 0.65 for LSTM and T-LSTM models to 0.69 for the T-LSTM model with one trainable variable parametric sigmoid.