Slide 1 – Title
Title: Machine Learning for 6-Hour-Ahead Sepsis Prediction: Logistic Regression vs Random Forest
“We worked on early prediction of sepsis in ICU patients. Our goal was to predict sepsis six hours
before clinical onset using only routine vital signs and basic patient information. We compare two
machine learning models: logistic regression and random forest, using data from the
PhysioNet/Computing in Cardiology 2019 Challenge, accessed via the Kaggle ‘Prediction of Sepsis’
dataset.”
Slide 2 – Study Aim & Research Question
Title: Study Aim & Research Question
“Our main goal is very focused: predict sepsis six hours before clinicians recognize it, using only
routine vital signs and simple patient/context information. We compare two models—logistic
regression and random forest—and we don’t just care about accuracy. We care about whether the
predictions would be clinically usable: how well they discriminate, how many alerts they generate,
and whether the behavior makes sense in a real ICU.”
Slide 3 – Why Early Sepsis Prediction?
Title: Why does early sepsis prediction matter?
“Sepsis is a major cause of ICU mortality, and every hour of delay in treatment worsens outcomes.
But ‘more alerts’ is not automatically better – false alarms cause alarm fatigue and misuse of limited
resources. That’s why sepsis prediction is a timing problem: we want alerts that are early enough to
act on, but accurate enough that clinicians can trust them.”
Slide 4 – Data & Features
Title: Data & Prediction Target
“We used the publicly available PhysioNet sepsis challenge dataset via Kaggle. Each row is one hour
of an ICU stay with a Patient_ID. We deliberately restricted ourselves to routine vitals and simple
context variables. That means heart rate, oxygen saturation, temperature, blood pressures,
respiratory rate, plus age, gender, and timing information such as ICU length of stay. For septic
patients, the label switches to 1 starting six hours before the recorded sepsis time; non-septic
patients stay at 0 throughout. So we are explicitly training a six-hour-ahead prediction model.”
Slide 5 – Methods: Preprocessing & Models
Title: Methods: Preprocessing & Models
“We first reconstructed each patient’s trajectory by sorting by Patient_ID and ICU hours. Missing
values were handled in two stages: forward-fill within each patient, then median imputation. We
removed extreme outliers based on interquartile ranges for main vitals. The data were then split into
80% training and 20% test, stratified by the sepsis label.
Sepsis is rare—about 1.5% of hourly rows—so we used SMOTEENN on the training set to oversample
septic cases and clean ambiguous ones. Finally, we trained two models: logistic regression, which is
linear and interpretable, and random forest, which captures nonlinear interactions by aggregating
many decision trees.”
Slide 6 – Results: LR vs RF (Test Set)
Title: Results (Test Set)
“On the held-out test set, sepsis remained rare at about 1.5% of hours.
Logistic regression reached an AUC of about 0.70 and recall around 0.58, meaning it caught many
positive hours, but its precision was extremely low. In other words, most predicted positives were
false alarms, and the F1-score was very poor.
Random forest performed better overall: AUC around 0.80, with a higher F1-score due to better
precision. However, even RF still produced quite a lot of false positives at the default 0.5 threshold.
So both models discriminate reasonably well, but in their current form they would generate a heavy
alert burden for clinicians.”
Slide 7 – Interpretation & Limitations
Title: Interpretation & Limitations
“Random forest clearly captures more structure in the data than logistic regression, which is
expected given the nonlinear physiology. LR’s advantage is interpretability and usually better
calibration.
There are several important limitations: we used a vitals-only feature set—no lab values—so
performance is probably a lower bound. We used SMOTEENN plus class weights, which can over-
emphasize the minority class and distort probabilities. We also split by rows, not strictly by patient,
so some patients may appear in both train and test. And we used the default 0.5 threshold, rather
than optimizing for a clinical utility function that trades off early detection vs false alarms. Finally,
missingness patterns and timing variables like ICULOS may reflect clinical behavior, not just
underlying physiology.”
Slide 8 – Conclusions & Future Work
Title: Conclusions & Next Steps
“In summary, machine learning models—especially random forests—can predict sepsis several hours
before clinical recognition using only vital signs and simple patient context. Logistic regression is
easier to explain; random forest is more accurate. However, if we deploy these models as-is, they
would cause too many false alarms.
Future work should focus on strictly patient-level splitting, better imbalance handling, explicit
calibration, and tuning thresholds based on a utility function or maximum acceptable alerts per
patient-day. Adding laboratory variables and exploring temporal deep learning models could further
improve early detection and clinical usefulness.”