Sepsis, a leading cause of hospital deaths, demands early prediction and diagnosis for effective mortality reduction. Current methods primarily rely on structured electronic medical record (EMR) data, neglecting the wealth of information within unstructured data like clinical notes (approximately 80% of EMR data). Early fluid resuscitation and antibiotic administration are critical within the first few hours of sepsis onset, highlighting the urgency of early detection. Delays in communication among healthcare professionals exacerbate this challenge. This study addresses this gap by developing an AI algorithm that leverages both structured and unstructured data (specifically, clinical notes) to improve the accuracy and lead time of sepsis prediction, aiming to provide clinicians with more time for effective treatment planning.

Existing sepsis prediction methods primarily utilize structured data from EMR systems. However, a significant portion of clinical data resides in unstructured formats, such as free-form text in clinical notes and images. These unstructured data contain rich clinical details not captured in structured fields. While some studies have incorporated text mining of clinical notes to enhance sepsis prediction, they often focus on identifying individual words, which can be limited by clinician-specific writing styles. This study extends prior work by using topic modeling to identify common themes within clinical notes, creating a more stable and generalizable model for sepsis prediction, enabling prediction up to 48 hours prior to sepsis onset.

The researchers developed the Sepsis Early Risk Assessment (SERA) algorithm, a topic-based, NLP-enabled AI algorithm that integrates the analysis of physicians' clinical notes with structured EMR data. The algorithm comprises two parts: a diagnosis algorithm to determine current sepsis status and an early prediction algorithm to assess the risk of sepsis within the next 4, 6, 12, 24, and 48 hours. The process involves the following steps: 1. **Data Acquisition and Preprocessing:** The study used data from a Singaporean hospital, including structured data (vital signs, lab results, treatments) and unstructured clinical notes from 482 sepsis patients in the training and validation sets, and 287 in the test set. The Synthetic Minority Over-sampling Technique (SMOTE) was used to balance the dataset due to the imbalanced nature of sepsis cases. 2. **Natural Language Processing (NLP):** Latent Dirichlet Allocation (LDA) topic modeling was applied to the clinical notes to identify 100 common topics, categorized into seven themes (admission, communication, lab tests, non-clinical status, social relationships, symptoms, and treatment). These topics were numerically weighted and combined with structured data. 3. **Machine Learning:** A voting ensemble machine learning algorithm (combining stochastic gradient descent (SGD)-based logistic regression and a random forest algorithm) was used for both diagnosis and early prediction. Gradient boosted trees (GBT) and DAGging were used for comparison. 4. **Model Evaluation:** The algorithm's performance was evaluated using metrics such as AUC, sensitivity, specificity, and positive predictive value (PPV) on both oversampled and non-oversampled datasets. The algorithm's performance was compared to physician predictions and existing scoring systems (SIRS, SOFA, MEWS). Finally, the impact of including unstructured clinical text on the model's performance was analyzed.

The SERA algorithm demonstrated high accuracy in both diagnosing current sepsis and predicting future onset. Key findings include: * **High Diagnostic Accuracy:** The diagnosis algorithm achieved an AUC of 0.94, sensitivity of 0.89, specificity of 0.85, and PPV of 0.95 in the test sample. * **Superior Early Prediction:** The early prediction algorithm showed high AUCs across various time horizons: 0.87 (48 hours), 0.90 (24 hours), 0.94 (12 hours), 0.92 (6 hours), and 0.92 (4 hours). The 12-hour prediction showed higher AUC, sensitivity, specificity, and PPV than previous studies. * **Improved Performance over Physicians:** The SERA algorithm outperformed physician predictions in terms of true positive rate (TPR), increasing early sepsis detection by 21-32% and reducing false positives by 7-17% across different time horizons. It also outperformed the accuracy of existing scoring systems (SOFA, MEWS, SIRS). * **Value of Unstructured Data:** Incorporating unstructured clinical notes significantly enhanced the algorithm's accuracy, particularly for predictions made 12-48 hours before sepsis onset. For the 12-hour prediction window, adding clinical text improved the AUC by 0.10-0.15, sensitivity by 0.07-0.13, and specificity by 0.08-0.14. * **Performance in Low Prevalence Settings:** Simulations showed that the algorithm maintains high sensitivity even in low sepsis prevalence environments, although PPV is naturally lower.

The SERA algorithm successfully addresses the challenge of early sepsis prediction by integrating both structured and unstructured clinical data. The algorithm's superior performance compared to physician predictions and existing scoring systems demonstrates its potential to improve sepsis management. The significant improvement in accuracy resulting from the inclusion of unstructured clinical notes highlights the untapped potential of this data source for clinical decision support. The ability to predict sepsis up to 48 hours in advance offers valuable time for clinicians to implement timely interventions, potentially reducing mortality. Further research should focus on real-world implementation and validation in diverse clinical settings to evaluate the algorithm's generalizability and scalability.

The SERA algorithm offers a significant advancement in early sepsis prediction and diagnosis. Its high accuracy, improved performance over existing methods, and the demonstrated value of incorporating unstructured data highlight its potential to improve patient outcomes. Future research should focus on broader validation across diverse populations and hospital settings, exploring integration with existing clinical workflows, and further enhancing the algorithm's interpretability to aid clinician trust and adoption.

The study's findings are based on data from a single hospital in Singapore, potentially limiting the generalizability of the results. Further validation in other healthcare systems with different patient populations and clinical practices is necessary. The study did not assess the impact of the algorithm on patient outcomes, such as mortality or length of hospital stay. Future research should conduct prospective studies to evaluate the clinical impact of using the SERA algorithm.