Disease Prediction from Text is the clinical NLP task of inferring likely diagnoses or disease risk from unstructured clinical narratives, patient-reported symptoms, and medical histories — enabling AI systems to predict clinical outcomes, generate differential diagnoses, flag high-risk patients, and identify undiagnosed conditions from the free-text content of electronic health records before formal diagnostic codes are assigned.
What Is Disease Prediction from Text?
- Task Scope: Ranges from binary disease classification (does this note suggest diabetes?) to multi-label multi-class diagnosis prediction across hundreds of ICD categories.
- Input: Chief complaint, history of present illness (HPI), past medical history, medications, lab results as text, nursing notes, clinical observation summaries.
- Output: Predicted ICD codes, disease probability scores, differential diagnosis list, or risk stratification label.
- Key Benchmarks: MIMIC-III (ICU discharge diagnosis prediction), n2c2 tasks (obesity and co-morbidity detection), eICU (multicenter ICU prediction), SemEval clinical NLP tasks.
The Clinical Prediction Task Types
Comorbidity Detection (NLP-based):
- Input: Discharge summary text.
- Output: Binary labels for 16 comorbidities (obesity, diabetes, hypertension, etc.).
- Benchmark: n2c2 2008 — 1,237 discharge summaries labeled for 15 obesity-related comorbidities.
Primary Diagnosis Prediction (ICD from text):
- Input: EHR notes before final coding.
- Output: Top-k predicted ICD-10 codes for the admission.
- Application: Pre-populate coding review queues; flag likely missed diagnoses.
Readmission Prediction:
- Input: Discharge summary text + structured data.
- Output: 30-day readmission risk binary classifier.
- Uses: Resource allocation, discharge planning, post-discharge follow-up intensity.
Mortality Prediction:
- Input: Clinical notes from first 24-48 hours of ICU admission.
- Output: In-hospital or 30-day mortality probability.
- Benchmark: MIMIC-III — state-of-the-art models achieve AUROC ~0.91 combining text + structured features.
Mental Health Screening:
- Input: Clinical note text or patient-reported questionnaire data.
- Output: PHQ-9 depression severity, suicide risk level, PTSD probability.
- Datasets: CLPSYCH shared tasks (depression and self-harm detection in social media and clinical notes).
Technical Approaches
TF-IDF + Classification: Simple bag-of-words baselines that perform surprisingly well on comorbidity detection (~85% micro-F1 on n2c2 2008).
ClinicalBERT / BioBERT:
- Fine-tuned on MIMIC-III for diagnosis prediction.
- Significant improvement over TF-IDF on rare comorbidities.
Hierarchical Models:
- For long documents (full discharge summary), hierarchically encode sections then aggregate.
- Section-level (admission note, progress notes, discharge summary) attention improves prediction by focusing on the most diagnostic text.
LLM-based with Structured Data:
- GPT-4 with patient timeline: structured lab values + unstructured notes → differential diagnosis + management chain.
- Achieves near-physician-level on curated cases; underperforms on complex multi-morbidity cases.
Performance Results
| Task | Best Model | Performance |
|------|-----------|------------|
| n2c2 2008 Comorbidity | ClinicalBERT | F1 ~93% |
| MIMIC-III 30-day readmission | BioBERT + structured | AUROC 0.736 |
| MIMIC-III in-hospital mortality | Multimodal LLM | AUROC 0.912 |
| MIMIC-III ICD prediction (top-50) | PLM-ICD | Micro-F1 0.798 |
Why Disease Prediction from Text Matters
- Undiagnosed Disease Detection: Clinical NLP can identify patterns suggesting undiagnosed conditions (undiagnosed diabetes in a patient presenting for an unrelated complaint) from note text before the physician has connected the dots.
- Sepsis Early Warning: Extracting fever, tachycardia, altered mental status, and bandemia from nursing notes before formal diagnosis flags sepsis 4-6 hours earlier than manual recognition.
- Oncology Surveillance: Cancer registry completion is ~60% accurate from structured data alone — text-based cancer identification from pathology reports and oncology notes captures the remainder.
- Preventive Care Gap Filling: Identifying patients with diabetes risk factors documented in notes but not yet in problem lists enables proactive screening outreach.
Disease Prediction from Text is the diagnostic intelligence layer of clinical AI — converting the rich narrative content of clinical documentation into actionable diagnostic signals that alert clinicians to urgent conditions, predict deterioration trajectories, and surface unrecognized disease burden hidden in the free text of electronic health records.