MIMIC-III Clinical Notes Overview

Updated 1 June 2026

MIMIC-III Clinical Notes are free-text records that capture granular, time-linked patient data from diverse clinical specialties.
They are preprocessed through PHI redaction, tokenization, and embedding techniques to enable effective feature extraction alongside structured data.
Machine learning approaches, including transformers and reinforcement learning, leverage these notes to enhance risk prediction, phenotyping, and automated coding.

MIMIC-III Clinical Notes are a central modality within the Medical Information Mart for Intensive Care III (MIMIC-III), an openly available database encompassing over 2 million free-text documents across nearly 60,000 critical-care admissions. These notes, written by clinicians of various specialties (physicians, nurses, radiologists, therapists), encode granular, time-linked patient information extending far beyond what is present in tabular physiological measurements or diagnosis codes. The multifaceted value of MIMIC-III notes spans NLP, clinical prediction, reinforcement learning for treatment policies, cohort phenotyping, and automated coding. Their integration in multimodal modeling pipelines has led to significant advances in risk prediction, patient stratification, and informatics-driven clinical research.

1. Structure, Content, and Taxonomy of MIMIC-III Clinical Notes

MIMIC-III clinical notes are stored in the NOTEEVENTS table and include multiple subtypes: progress notes, nursing entries, radiology reports, discharge summaries, ECG and echo interpretations, and various administrative or consult notes. Only a minority of hospital admissions (<9%) contain SOAP-structured (Subjective, Objective, Assessment, Plan) progress notes, whereas nearly all contain nursing and radiology notes (Keerthana et al., 18 Jul 2025). Note structure is variable—some are sectioned by explicit headers (e.g., “History of Present Illness”), while others are free-form texts with inconsistent labeling or formatting. This heterogeneity necessitates careful curation; DENSE reclassifies the entire NOTEEVENTS corpus into 16 coherent types to enable structured temporal modeling (Keerthana et al., 18 Jul 2025).

Notes are time-stamped and can be linked to specific hospital admissions (HADM_ID) and ICU stays. For longitudinal applications, progress notes are often chronologically concatenated or pivoted using “visit pivoting” to create admission-centric data structures (Keerthana et al., 18 Jul 2025).

2. Preprocessing, Representation, and Feature Extraction

Preprocessing of MIMIC-III notes encompasses PHI redaction, lowercasing, tokenization (whitespace, punctuation), and variable use of stop-word removal and stemming. Lemmatization, spelling correction, or abbreviation expansion are typically omitted; clinical concepts are sometimes processed at the token level, but advanced pipelines may extract UMLS Concept Unique Identifiers (CUIs) or map spans to phenotype ontologies (Shin et al., 2021, Zhang et al., 2021, Li et al., 2018).

Feature representations include:

Bag-of-words and TF–IDF vectors: Count-based, often high-dimensional (e.g., 7,248 unigrams after filtering for tokens in ≥10 patient-note documents) (Shin et al., 2021).
Distributed embeddings: Pre-trained on MIMIC-III or external clinical text using Word2Vec, FastText, or contextual encoders (e.g., ClinicalBERT). In some pipelines, embeddings are average-pooled per note or concatenated with structured features (Hashir et al., 2019, Kemp et al., 2019, King et al., 2023).
Contextual transformers: ClinicalBERT, standard BERT, and specialized LLMs such as Med42-v2-70B are increasingly used for tokenization, passage encoding, and for generating summaries or explanations (Battula et al., 2024, Grundmann et al., 2021).
Ontology-driven feature extraction: Named Entity Recognition (NER) and concept linking tools such as MetaMap or SemEHR map text to UMLS/ORDO codes, improving rare-disease and phenotype identification (Li et al., 2018, Dong et al., 2022, Zhang et al., 2021).

Notably, section segmentation (extracting “Assessment,” “Plan,” etc.) is critical for certain tasks such as note synthesis and answer retrieval (Keerthana et al., 18 Jul 2025, Grundmann et al., 2021), but is not universally applied.

3. Machine Learning and Deep Learning Integration

Predictive Modeling

Early Mortality and Morbidity Prediction: Combined models incorporating both structured (vitals, labs, demographics) and unstructured (TF–IDF, embeddings, notes) features consistently outperform single-modality models:

For sepsis mortality, L2-regularized logistic regression fusing 44 structured features with 7,248 TF–IDF text features achieves AUC=0.842, F₁=0.512, exceeding either input alone (Shin et al., 2021).
Hierarchical CNN–RNN architectures, operating on minimally cleaned free-text note sequences, yield AUROC up to 0.9023 on in-hospital mortality, outperforming time-series-only and bag-of-words models (Hashir et al., 2019).

Phenotyping and Cohort Extraction: Transformer-based BiLSTM-CNN hybrids and ontology-driven pipelines can classify complex phenotypes (e.g., advanced cancer, psychiatric disorders) with F1-micro up to 0.636 and F1-macro up to 0.614 (Khalafi et al., 2021). Self-supervised contextual synonym detection achieves high AUC in phenotype-driven risk modeling by mapping text spans to the Human Phenotype Ontology and propagating persistent vs. transient features (Zhang et al., 2021).

Automated Coding: State-of-the-art transformer models (BERT- or AWD-LSTM-based) fine-tuned on MIMIC-III notes achieve 93.76% accuracy and F1=92.24% for top-50 code multi-label classification, with AUC=91% on test (Singh et al., 2020, Nuthakki et al., 2019, Huang et al., 2018). Models pretraining contextual LMs on MIMIC-III followed by hierarchical patient-level inference demonstrate further gains in recall and diagnostic accuracy (Kemp et al., 2019).

4. Multimodal Modeling and Longitudinal Synthesis

Contemporary research leverages joint modeling of notes and structured data using multimodal representation learning, self-supervised alignment, and LLM-augmented document synthesis:

Multi-view architectures (e.g., multi-representational models) concatenating LSTM-based structured branch, BERT-embedded raw notes, and LLM-generated expert summaries achieve AUROC=0.8955, AUPRC=0.6156, with marked improvements in fairness across demographic subgroups (Battula et al., 2024).
Contrastive and masked-token pretraining aligns time-series and note encodings, boosting AUC-ROC for mortality by 0.17 at 1% label availability—critical for low-supervision applications (King et al., 2023).
LLM-driven note synthesis (DENSE) reconstructs missing SOAP progress notes by temporally aligning, semantically retrieving, and generating notes with LLM prompting. The method achieves a temporal alignment ratio of 1.089, indicating higher longitudinal coherence than real progress notes (Keerthana et al., 18 Jul 2025).

Reinforcement Learning Integration

Offline RL for sepsis management fuses LLM-derived note embeddings and high-frequency structured data through gated fusion and cross-modal attention, producing enriched state vectors. Multimodal RL increases estimated survival rates and policy value by 15–24% (OPERA, DR), with FQE improvement from 0.622 to 1.382 on MIMIC-III (Lim et al., 11 Aug 2025).
Modeling informative missingness in notes availability (explicitly modeling text-missingness, documentation gaps) further drives gains in offline policy evaluation and mortality prediction (FQE 0.679 vs. 0.528; AUROC 0.886 vs. 0.844) (Liang et al., 23 Apr 2026).

5. Interpretation, Fairness, and Evaluation Practices

Interpretability: SHAP (Shapley Additive Explanations) analyses show that automatically extracted phenotypes (e.g., pain, constitutional symptoms) are among the most influential cohort-level features for mortality prediction, while patient-level SHAP scores reveal time-localized impact of new note content on risk trajectories (Zhang et al., 2021). Conventional linear models (logistic regression, SVM) retain favor for their feature interpretability in high-dimensional token spaces (Shin et al., 2021).

Equity: Multi-representational LLM-augmented models demonstrate more equitable performance gains across racial and ethnic subgroups, substantially increasing AUROC and AUPRC for underrepresented populations while reducing model bias (Battula et al., 2024).

Silver-Standard Labeling and Evaluation: MIMIC-III “gold-standard” ICD-9 code assignments exhibit systematic undercoding—raw assignments omit up to 35% of true diagnoses for frequent conditions. Use of section-based NER and manual validation to construct silver-standard labels is recommended to mitigate label noise and improve robust evaluation (Searle et al., 2020).

Quantitative Performance Table

Task	Model/Data Modality	Best Reported Metric(s)	Reference
30-day Sepsis Mortality	L2-LR, text+structured	AUC=0.842, F1=0.512	(Shin et al., 2021)
ICU Mortality (notes+TS, HCR)	MM-HCR (text+TS)	AUROC=0.9023	(Hashir et al., 2019)
Phenotyping (10 traits)	Hybrid BiGRU-CNN	F1-micro=0.636	(Khalafi et al., 2021)
ICD-9 Coding (top-50)	BERT, multi-label	Acc=93.76%, F1=92.24%	(Singh et al., 2020)
Longitudinal Synth. (temporal align.)	DENSE	TAR=1.089	(Keerthana et al., 18 Jul 2025)
RL Action-value (FQE, sepsis)	Multimodal RL	FQE=1.382	(Lim et al., 11 Aug 2025)
Mortality Prediction, 48h, multi-rep LLM	LLM summary+notes+TS	AUROC=0.8955, Δ+7.6%	(Battula et al., 2024)

6. Limitations, Open Problems, and Future Directions

Data Sparsity and Heterogeneity: Only a small fraction of stays include structured progress notes, with many studies relying on noisy or semi-structured textual artifacts. Ongoing research on taxonomy normalization, section segmentation, and meticulous curation is necessary to enable robust cross-cohort modeling (Keerthana et al., 18 Jul 2025).
Label Noise in Gold Standards: Legacy ICD-9 coding in MIMIC-III shows substantial undercoverage of true diagnoses. Supplementing analysis with silver-standard datasets and built-in adjudication pipelines is recommended (Searle et al., 2020).
External and Prospective Validation: Most pipelines are benchmarked solely on MIMIC-III. Generalizability to MIMIC-IV and real-world EHRs with different note structures, coding practices, and population demographics remains an essential open area (Keerthana et al., 18 Jul 2025, Battula et al., 2024).
LLM Hallucination and Trust: LLM-generated summaries and note syntheses improve performance but risk propagation of hallucinations or implicit bias. Human-in-the-loop validation and scalable, transparent reliability estimation are current research imperatives (Battula et al., 2024, Keerthana et al., 18 Jul 2025).
Representational and Methodological Innovation: Opportunities for advancement include hierarchical and event-based document modeling, robust multi-task contrastive pretraining across modalities, domain-adapted LLMs, and the integration of richer ontologies and structured knowledge (e.g., UMLS, HPO, SNOMED CT) (Zhang et al., 2021, Dong et al., 2022, King et al., 2023).

7. Research Impact and Practical Applications

MIMIC-III clinical notes, when systematically processed and fused with other EHR modalities, unlock a spectrum of clinical informatics tasks: high-resolution risk prediction, end-to-end diagnostic and procedure coding, automated phenotype annotation, time-aware note synthesis, and deep offline RL for sequential decision making. The current state-of-the-art demonstrates that even minimal NLP pipelines, when coupled to robust embedding and learning architectures, yield substantial gains on all major critical-care benchmarks. Broader adoption hinges on advances in scalability, explainability, data fidelity, and regulatory-compliant deployment.