NLP and Machine Learning for Early Sepsis Detection in Neonatal Care

wroclaw_medical_university · Healthcare · year n/a · source Agentic: no · promotional bias: low · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Neonatal clinical decision support

Wroclaw Medical University needed to detect early signs of sepsis in newborns by combining unstructured EHR text with real-time clinical signals.

Constraints:

Sepsis is life-threatening and requires rapid diagnosis.
Relevant data includes vital signs, lab results, and clinical observations.
Electronic health records contain unstructured medical data.
Clinical use requires sensitivity, specificity, and avoidance of unnecessary antibiotics.

Architecture

The project combines NLP extraction from EHR text with machine learning over structured clinical signals to support early sepsis detection in neonatal intensive care units.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  electronic-health-records["Electronic Health Records"]
  nlp-extraction["NLP Extraction"]
  vital-signs-and-lab-results["Vital Signs and Lab Results"]
  sepsis-detection-model["Sepsis Detection Model"]
  clinical-decision-support["Clinical Decision Support"]
  electronic-health-records --> nlp-extraction
  nlp-extraction --> sepsis-detection-model
  vital-signs-and-lab-results --> sepsis-detection-model
  sepsis-detection-model --> clinical-decision-support

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Electronic Health Records`	Provide unstructured neonatal medical data.	`NLP Extraction`
`NLP Extraction`	Structures clinical observations from unstructured text.	`Sepsis Detection Model`
`Vital Signs and Lab Results`	Provide real-time structured clinical signals.	`Sepsis Detection Model`
`Sepsis Detection Model`	Detects early patterns of neonatal sepsis.	`Clinical Decision Support`
`Clinical Decision Support`	Supports faster diagnosis and antibiotic decisions.	—

Data flow

Unstructured EHR notes are processed with NLP, vital signs and lab results are combined with extracted observations, machine learning models detect early sepsis patterns, and the clinical decision support system alerts or informs neonatal care teams.

“uses natural language processing to extract and structure unstructured medical data from electronic health records, combined with machine learning models to detect early signs of sepsis in newborns.”summary, stated

“by analyzing vital signs, lab results, and clinical observations in real-time.”summary, stated

Why this architecture

Combining NLP with structured real-time signals is necessary because sepsis evidence sits across notes, vital signs, lab results, and observations.

“extract and structure unstructured medical data from electronic health records”summary, stated

“analyzing vital signs, lab results, and clinical observations in real-time.”summary, stated

Models used: Natural language processing models · Machine learning sepsis detection models

Scale

Sepsis causes 3 million deaths annually.
“sepsis—a life-threatening condition that causes 3 million deaths annually”summary, stated
Similar systems cited could reduce diagnosis time from 24 hours to 2 hours.
“reduce diagnosis time from 24 hours to 2 hours”summary, stated

Alternatives considered (roads not taken)

Manual clinical recognition alone

Rejected because: Rapid diagnosis is critical and signals are spread across real-time data and EHR text.

“requires rapid diagnosis”summary, stated

Structured data only

Rejected because: Important clinical observations are embedded in unstructured EHR data.

“extract and structure unstructured medical data from electronic health records”summary, stated

Patterns

`safety-policy-enforcement` (clinical-safety)

The decision support system targets high-sensitivity early detection while also reducing unnecessary antibiotics.

“high sensitivity and specificity, while also potentially reducing unnecessary antibiotic use through pattern recognition.”summary, stated

`eval-driven-development` (evaluation)

The case frames value through sensitivity, specificity, diagnosis time, and antibiotic-use outcomes.

“with high sensitivity and specificity”summary, stated

Tradeoffs

Use combined NLP and ML

Chose Text extraction plus structured signal modeling over Single-source detection.

Rationale: Sepsis evidence spans notes, vital signs, lab results, and observations.

“uses natural language processing to extract and structure unstructured medical data from electronic health records, combined with machine learning models”summary, stated

Optimize for earlier detection

Chose Real-time clinical decision support over Delayed retrospective review.

Rationale: Sepsis requires rapid diagnosis and intervention.

“requires rapid diagnosis”summary, stated

Gotchas (what bit them)

Early sepsis signals are distributed across text and structured clinical feeds.

Root cause: EHR notes, vital signs, labs, and observations each carry partial evidence.

Resolution: Combine NLP over EHR text with ML over real-time clinical signals.

“analyzing vital signs, lab results, and clinical observations in real-time.”summary, stated

False positives can drive unnecessary antibiotic use.

Root cause: Clinical alerting must balance rapid diagnosis with treatment burden.

Resolution: Use pattern recognition with sensitivity and specificity targets.

“potentially reducing unnecessary antibiotic use through pattern recognition.”summary, stated

Evaluation

The presentation cites expected clinical metrics from similar systems, including diagnosis-time reduction, sensitivity, specificity, and antibiotic-use impact.

Metrics: Diagnosis time · Sensitivity · Specificity · Unnecessary antibiotic use

“reduce diagnosis time from 24 hours to 2 hours, with high sensitivity and specificity”summary, stated

Outcomes

The project is developing a neonatal ICU decision-support system.
Similar systems cited reduce diagnosis time from 24 hours to 2 hours.
The approach may reduce unnecessary antibiotic use.

The case cites 3 million annual sepsis deaths and neonatal ICU deployment context, but not hospital rollout count.

“clinical decision support system for neonatal intensive care units”summary, stated