Abstract:
Sepsis is the body’s abnormal and dysregulated response to an infection due to
septicaemia. It causes multiple organ damage, and eventually, the patient dies. The
worldwide mortality ratio of sepsis is exceptionally high, with an estimated 11 million
deaths according to 2017 global sepsis statistics. In Pakistan, an estimated 60-80% of
intensive care unit (ICU) deaths are due to sepsis, which might reach 90% soon. Due to
limited resources and inflation, early sepsis detection is imperative to lower mortality.
Several machine learning-based sepsis prediction tools have been developed, and many
studies have been conducted for sepsis prediction. However, these tools cannot predict
sepsis as it is a time-series problem but treat it as a binary classification problem. Deep-
learning (DL) algorithm-based methods can better deal with the time-series data due to
their robustness, allowing better insights into the data and performance. Therefore, in
this study, a novel DL-based approach is opted to forecast the sepsis mortality risk in
ICU patients. MissForest and Last Observation Carried Forward (LOCF)-zero (FFILL-
0) imputation methods were used to impute Not a Number (NaN) values (missing
values) in the data, and the patient data was converted to fixed-length tensors, which
were then used for model training and evaluation. Among the DL algorithms, Long
short-term memory (LSTM) and Gated recurrent units (GRUs) were selected for model
building. Finally, four models were trained on MissForest, and FFILL-0 imputed data,
and to check the effectiveness, the models were evaluated on the hold-out datasets and
the Area Under the Receiver Operating Characteristic curve (AUROC) was calculated
for each model. LSTM model outperformed GRU, and the highest AUROC achieved
in this study was 0.758. In short, DL algorithms can accurately forecast sepsis risk in
ICU patients and can help reduce Intensive Care Unit Length of Stay (ICULOS) and
sepsis mortality risk.
