Automated Detection of Dosing Errors in Clinical Trial Narratives: A Multi-Modal Feature Engineering Approach with LightGBM

Mohammad AL-Smadi

arxiv: 2604.19759 · v1 · submitted 2026-03-25 · 💻 cs.AI · cs.CL

Automated Detection of Dosing Errors in Clinical Trial Narratives: A Multi-Modal Feature Engineering Approach with LightGBM

Mohammad AL-Smadi This is my paper

Pith reviewed 2026-05-15 00:42 UTC · model grok-4.3

classification 💻 cs.AI cs.CL

keywords dosing error detectionclinical trial narrativesLightGBMmulti-modal featuresclass imbalanceROC-AUCfeature selectionnatural language processing

0 comments

The pith

A LightGBM classifier with 3,451 multi-modal features detects dosing errors in clinical trial narratives at 0.8725 test ROC-AUC despite only 4.9 percent positive cases.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper builds an automated detector for dosing errors inside unstructured clinical trial text by feeding a gradient-boosting model thousands of engineered features drawn from nine separate narrative fields. Traditional bag-of-words measures, character n-grams, sentence embeddings, and transformer scores from BiomedBERT and DeBERTa are combined into one 3,451-dimensional representation for each of 42,112 samples. On a benchmark set with extreme imbalance the resulting LightGBM ensemble reaches 0.8725 ROC-AUC after five-fold averaging, and ablation experiments show that dropping the sentence embeddings hurts performance most. Feature-selection experiments further reveal that keeping only the top 500–1,000 features actually improves the score by cutting noise. The work matters because manual review of these long narratives is slow and fallible, so a reliable automated screen could reduce missed errors that affect patient safety and trial validity.

Core claim

The central claim is that a LightGBM model trained on a carefully engineered mix of sparse lexical, dense semantic, and domain-specific medical features extracted from nine complementary text fields can reliably flag dosing errors even when positive examples constitute less than five percent of the data, reaching 0.8725 test ROC-AUC while systematic ablations and feature-efficiency tests confirm that sentence embeddings are indispensable and that aggressive feature selection improves rather than harms accuracy.

What carries the argument

The 3,451-dimensional multi-modal feature vector combining TF-IDF, character n-grams, all-MiniLM-L6v2 embeddings, BiomedBERT and DeBERTa-v3 scores, plus hand-crafted medical patterns, used to train a LightGBM classifier.

If this is right

Reducing the feature set to the top 500–1000 dimensions raises AUC above the full 3,451-feature baseline by removing noise.
Sentence embeddings account for the single largest performance drop when removed, even though they contribute only 37 percent of total importance.
Sparse lexical features remain complementary to dense transformer scores for clinical text tasks under severe imbalance.
Five-fold ensemble averaging stabilizes predictions enough to reach 0.8725 test AUC from 0.8833 cross-validation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same pipeline could be retrained periodically on newly completed trials to keep the detector current with evolving protocol language.
Similar multi-modal feature sets might transfer to detection of other protocol violations such as eligibility errors or adverse-event misreporting.
The observed benefit of aggressive feature selection suggests the method could run on modest hardware in hospital data warehouses without retraining from scratch each time.

Load-bearing premise

The nine text fields and the features extracted from them capture essentially all relevant signals of dosing errors without systematic missing context or label noise in the CT-DEB collection.

What would settle it

Retraining and testing the identical pipeline on an independently annotated collection of clinical trial narratives that was never seen during feature design would show whether the 0.87 AUC holds or collapses.

read the original abstract

Clinical trials require strict adherence to medication protocols, yet dosing errors remain a persistent challenge affecting patient safety and trial integrity. We present an automated system for detecting dosing errors in unstructured clinical trial narratives using gradient boosting with comprehensive multi-modal feature engineering. Our approach combines 3,451 features spanning traditional NLP (TF-IDF, character n-grams), dense semantic embeddings (all-MiniLM-L6v2), domain-specific medical patterns, and transformer-based scores (BiomedBERT, DeBERTa-v3), used to train a LightGBM model. Features are extracted from nine complementary text fields (median 5,400 characters per sample) ensuring complete coverage across all 42,112 clinical trial narratives. On the CT-DEB benchmark dataset with severe class imbalance (4.9% positive rate), we achieve 0.8725 test ROC-AUC through 5-fold ensemble averaging (cross-validation: 0.8833 + 0.0091 AUC). Systematic ablation studies reveal that removing sentence embeddings causes the largest performance degradation (2.39%), demonstrating their critical role despite contributing only 37.07% of total feature importance. Feature efficiency analysis demonstrates that selecting the top 500-1000 features yields optimal performance (0.886-0.887 AUC), outperforming the full 3,451-feature set (0.879 AUC) through effective noise reduction. Our findings highlight the importance of feature selection as a regularization technique and demonstrate that sparse lexical features remain complementary to dense representations for specialized clinical text classification under severe class imbalance.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This applies standard LightGBM plus multi-modal features to dosing error detection and gets usable AUC numbers with some clean ablation work, but missing baselines and label details limit how far the claims travel.

read the letter

The core contribution is a practical pipeline that pulls 3,451 features from nine text fields in 42k clinical trial narratives and trains LightGBM to catch dosing errors. On the CT-DEB set with 4.9% positives it reaches 0.8725 test AUC via 5-fold averaging, and the ablations show sentence embeddings matter most while top-500-1000 features actually beat the full set by cutting noise. That feature-efficiency result is the most useful takeaway for anyone doing similar clinical text work. The nine-field coverage and the mix of TF-IDF, embeddings, and BiomedBERT scores look thorough for the task. The paper stays grounded in concrete numbers and reports the cross-validation spread, which is better than many applied ML abstracts. The main gaps are straightforward. No baseline models or prior clinical-text classifiers are shown, so it is hard to judge whether the multi-modal engineering delivers a real lift or just matches what simpler approaches already achieve. The label side is thin: nothing is said about how positives were identified, inter-annotator agreement, or adjudication rules, which leaves open the possibility that the AUC reflects dataset artifacts more than true signal. No external validation set is mentioned either. This is for readers who build or evaluate clinical NLP tools for safety monitoring. They will get value from the ablation tables and the feature-selection lesson. It is not a theoretical paper and does not claim broad impact, but the results are internally consistent enough that a serious editor should send it to referees with requests for baselines and label provenance.

Referee Report

3 major / 2 minor

Summary. The manuscript presents a LightGBM-based classifier for automated detection of dosing errors in clinical trial narratives. It extracts 3,451 multi-modal features (TF-IDF, character n-grams, all-MiniLM-L6v2 embeddings, BiomedBERT/DeBERTa-v3 scores, and domain patterns) from nine text fields across the full CT-DEB dataset of 42,112 samples (4.9% positive class) and reports a test ROC-AUC of 0.8725 via 5-fold ensemble averaging (CV AUC 0.8833 ± 0.0091). Ablation studies and feature-selection experiments are included, showing that sentence embeddings contribute critically and that top-500–1000 features outperform the full set.

Significance. If the labels prove reliable, the work offers a concrete demonstration of multi-modal feature engineering for severely imbalanced clinical-text classification, with reproducible metrics, ablation results, and evidence that feature selection acts as effective regularization. The approach could support safety monitoring in trials, but its significance is limited by the lack of external validation and unverified label provenance, which directly affects whether the 0.8725 AUC reflects real-world utility rather than an upper bound.

major comments (3)

[Data section] Data section (or Methods, dataset description): No information is provided on label provenance, including the expert review process, inter-annotator agreement, or adjudication rules used to create the CT-DEB positive/negative labels. Because the central performance claim (0.8725 test ROC-AUC) rests on the assumption that the nine text fields plus engineered features fully capture dosing-error signals with negligible label noise, this omission is load-bearing and must be addressed before the result can be interpreted as reliable.
[Results section] Results section: The manuscript reports no comparisons against prior clinical-text baselines (e.g., simpler TF-IDF + logistic regression or existing medical NLP models on similar tasks). Without these, it is impossible to determine whether the 3,451-feature LightGBM pipeline provides meaningful improvement over established methods, weakening the claim that the multi-modal approach is superior under class imbalance.
[Ablation studies] Ablation and feature-importance analysis: While the 2.39% drop from removing sentence embeddings is reported, the paper does not state whether this difference is statistically significant across folds or whether the importance ranking (embeddings at 37.07%) was computed on the training or test set; either clarification is needed to support the conclusion that embeddings are “critical” despite lower importance share.

minor comments (2)

[Abstract] Abstract and Methods: The exact breakdown of the 3,451 features (counts per modality) and the precise definition of “domain-specific medical patterns” should be tabulated for reproducibility.
[Introduction] The manuscript should cite prior work on clinical-trial text classification and dosing-error detection to situate the contribution.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for the constructive and detailed feedback on our manuscript. We address each major comment below and indicate the revisions we will make to the next version of the paper.

read point-by-point responses

Referee: [Data section] Data section (or Methods, dataset description): No information is provided on label provenance, including the expert review process, inter-annotator agreement, or adjudication rules used to create the CT-DEB positive/negative labels. Because the central performance claim (0.8725 test ROC-AUC) rests on the assumption that the nine text fields plus engineered features fully capture dosing-error signals with negligible label noise, this omission is load-bearing and must be addressed before the result can be interpreted as reliable.

Authors: We agree that label provenance is essential for interpreting the reliability of the reported AUC. The CT-DEB dataset was used as a pre-existing benchmark collection, and detailed annotation metadata (expert review process, inter-annotator agreement, and adjudication rules) are not available in the dataset documentation or to the authors. In the revised manuscript we will expand the Data section to explicitly describe the known characteristics of the labels, state the source of the benchmark, and add a limitations paragraph discussing the implications of unknown label noise. revision: partial
Referee: [Results section] Results section: The manuscript reports no comparisons against prior clinical-text baselines (e.g., simpler TF-IDF + logistic regression or existing medical NLP models on similar tasks). Without these, it is impossible to determine whether the 3,451-feature LightGBM pipeline provides meaningful improvement over established methods, weakening the claim that the multi-modal approach is superior under class imbalance.

Authors: We concur that baseline comparisons are necessary to substantiate the contribution of the multi-modal pipeline. In the revised Results section we will add performance numbers for two standard baselines on the identical CT-DEB split: (1) TF-IDF features with logistic regression and (2) a fine-tuned all-MiniLM-L6v2 classifier. These additions will allow direct assessment of whether the 3,451-feature LightGBM model yields meaningful gains under the reported class imbalance. revision: yes
Referee: [Ablation studies] Ablation and feature-importance analysis: While the 2.39% drop from removing sentence embeddings is reported, the paper does not state whether this difference is statistically significant across folds or whether the importance ranking (embeddings at 37.07%) was computed on the training or test set; either clarification is needed to support the conclusion that embeddings are “critical” despite lower importance share.

Authors: We thank the referee for highlighting this ambiguity. Feature importance (including the 37.07% share attributed to embeddings) was obtained from LightGBM’s gain metric averaged over the training folds; no test-set information was used. The 2.39% AUC degradation was observed in every one of the five folds. In the revision we will explicitly document these computation details and add a paired t-test across the per-fold AUC values to evaluate statistical significance of the embedding ablation. revision: yes

standing simulated objections not resolved

Specific details on the expert review process, inter-annotator agreement, and adjudication rules used to generate the CT-DEB labels, which are not documented in the benchmark and therefore cannot be supplied by the authors.

Circularity Check

0 steps flagged

No circularity: standard supervised ML pipeline with held-out evaluation

full rationale

The paper describes a conventional feature-engineering + LightGBM classification pipeline on a labeled dataset (CT-DEB). Performance is measured via 5-fold cross-validation and held-out test ROC-AUC; no equation, prediction, or central claim reduces to a fitted parameter by construction, nor does any load-bearing step rely on self-citation or self-definition. The nine text fields and 3,451 features are inputs to the model, not outputs derived from the reported AUC. Label provenance and field completeness are external assumptions (correctness risk), not circularity in the derivation chain.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The work rests on standard supervised classification assumptions and the representativeness of the CT-DEB labels; no new entities or free parameters beyond routine hyperparameter tuning are introduced.

free parameters (2)

LightGBM hyperparameters
Tuned on validation data to achieve reported AUC; exact values not stated in abstract.
Top feature count
Selected post-hoc as 500-1000 for optimal performance.

axioms (1)

domain assumption CT-DEB dataset labels accurately reflect true dosing errors
Required for supervised training and AUC evaluation.

pith-pipeline@v0.9.0 · 5584 in / 1217 out tokens · 51023 ms · 2026-05-15T00:42:42.165216+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

20 extracted references · 20 canonical work pages · 1 internal anchor

[1]

dose adjusted per investigator discretion

Introduction Clinical trials are fundamental to pharmaceutical development and medical advancement, requir- ing strict adherence to pre-defined protocols spec- ifying medication dosing, timing, and administra- tion routes. Dosing errors—deviations from these protocols—pose significant risks to patient safety and trial validity (ICH Expert Working Group, 2...

work page 2001
[2]

A 3,451-dimensional feature space combin- ing traditional NLP (Term Frequency–Inverse Document Frequency (TF-IDF), character n- grams), sentence embeddings, medical pat- terns, and transformer-based scores

work page
[3]

Systematicanalysisshowingsentenceembed- dings and lexical features are complementary, while transformer scores underperform

work page
[4]

Optuna-optimized 5-fold ensemble achieving 0.8833 ± 0.0091 CV AUC with minimal overfit- ting (0.69% out-of-fold(OOF)-test gap)

work page
[5]

Automated Detection of Dosing Errors in Clinical Trial Narratives: A Multi-Modal Feature Engineering Approach with LightGBM

Feature selection improves performance: 500- 1000 features (14-29%) achieve 0.886-0.887 AUC, outperforming the full baseline (0.879) through noise reduction. arXiv:2604.19759v1 [cs.AI] 25 Mar 2026

work page internal anchor Pith review Pith/arXiv arXiv 2026
[6]

Test ROC-AUC of 0.8725 on CT-DEB with threshold-adjustable recall (26-60%) for flexi- ble deployment

work page
[7]

Clinical NLP and Information Extraction Clinical NLP has evolved significantly over the past two decades

Related Work 2.1. Clinical NLP and Information Extraction Clinical NLP has evolved significantly over the past two decades. Early systems like cTAKES (Savova et al., 2010) and MetaMap (Aronson and Lang,

work page 2010
[8]

provided rule-based approaches for med- ical concept extraction and normalization. The i2b2 shared tasks (Uzuner et al., 2011) established benchmarks for concept extraction, assertion de- tection, and relation extraction from clinical texts, demonstrating that machine learning approaches could achieve substantial performance gains. More recent work has fo...

work page 2011
[9]

Dataset and Task 3.1. CT-DEB Benchmark We utilize the CT-DEB (Clinical Trial Dosing Error Benchmark) dataset (Hêche et al., 2026), specif- ically designed for evaluating automated dosing error detection systems. The dataset comprises clinical trial narratives describing medication ad- ministration across various therapeutic areas, pro- tocols, and clinica...

work page 2026
[10]

adverse event

Methodology 4.1. Data Preparation and Feature Engineering Pipeline Our feature engineering pipeline transforms raw clinical narratives into a 3,451-dimensional feature vector combining multiple representation types. Ta- ble 2 provides an overview of extracted features. Clinical trial registry data exhibits inherent spar- sity across structured fields. To ...

work page 1988
[11]

Train LightGBM on the remaining 4 folds (80% of data) using weighted binary cross-entropy: L=− 1 N NX i=1 wi [yi log(ˆyi) + (1−y i) log(1−ˆyi)] (2) where wi = 20.87forpositiveexamples, wi = 1 for negative

work page
[12]

Generate OOF predictions on foldk (the held- out 20%)

work page
[13]

Ensemble Prediction- For test inference, all 5 fold models generate predictions, which are aver- aged: ˆyensemble(x) = 1 5 5X k=1 ˆyk(x)(3) ct-dosing-errors-benchmark

Save the trained model for later ensembling Out-of-Fold Validation- Concatenating predic- tions from all 5 folds yields complete OOF pre- dictions across the training set, providing an un- biased performance estimate: 0.8833 ± 0.0091 ROC-AUC. Ensemble Prediction- For test inference, all 5 fold models generate predictions, which are aver- aged: ˆyensemble(...

work page
[14]

reduced”, “discontinued

Results 5.1. Overall Performance We employ 5-fold stratified cross-validation with ensemble averaging. Table 5 presents test set per- formance. The ROC-AUC of 0.8725 indicates excellent dis- criminativeability. Theensembledemonstratessta- ble performance with mean cross-validation AUC of 0.8833 ± 0.0091 across folds, and OOF AUC of 0.8794, showing minimal...

work page
[15]

This variation reflects different train- ing/validation partitions while maintaining robust performance

when using early stopping with 200-iteration patience. This variation reflects different train- ing/validation partitions while maintaining robust performance. Cross-validation stability: Mean fold AUC of 0.883 ± 0.009 demonstrates low variance across datasplits,withindividualfoldperformanceranging from0.869to0.894(2.5%range). Thisconsistency indicates ro...

work page
[16]

Most predictive value is concentrated in a small-to-moderate subset of features (500-1000), with the remaining features introducing more noise than signal

Discussion The findings indicate that accurate detection of dos- ingdeviationsinclinicalnarrativesreliesoncombin- ing sparse lexical features with contextual embed- dings,aseachcapturesdistinctbutcomplementary signals. Most predictive value is concentrated in a small-to-moderate subset of features (500-1000), with the remaining features introducing more n...

work page
[17]

Conclusion We present an automated system for detecting dosing errors in clinical trial narratives, achieving 0.8725 test ROC-AUC through 5-fold ensemble learning with comprehensive multi-modal feature engineering. Our approach combines 3,451 fea- tures spanning traditional NLP (TF-IDF, character n-grams),densesemanticembeddings(all-MiniLM- L6-v2), handcr...

work page
[18]

References Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. 2019. Op- tuna: A next-generation hyperparameter opti- mization framework. InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 2623–2631. Emily Alsentzer, John R. Murphy, Willie Boag, Wei-Hung Weng, Di Jin, Trist...

work page 2019
[19]

clinical practice guidelines

Domain-specific language model pretrain- ing for biomedical natural language processing. ACM Transactions on Computing for Healthcare, 3(1):1–23. Rave Harpaz, William DuMouchel, Nigam H. Shah, David Madigan, Patrick Ryan, and Carol Fried- man.2012. Noveldata-miningmethodologiesfor adverse drug event discovery and analysis.Clin- ical Pharmacology & Therape...

work page arXiv 2012
[20]

Özlem Uzuner, Brett R

Enhancing clinical concept extraction with contextual embeddings.Journal of the American Medical Informatics Association, 26(11):1297– 1304. Özlem Uzuner, Brett R. South, Shuying Shen, and Scott L. DuVall. 2011. 2010 i2b2/va challenge on concepts, assertions, and relations in clinical text.JournaloftheAmericanMedicalInformatics Association, 18(5):552–556....

work page 2011

[1] [1]

dose adjusted per investigator discretion

Introduction Clinical trials are fundamental to pharmaceutical development and medical advancement, requir- ing strict adherence to pre-defined protocols spec- ifying medication dosing, timing, and administra- tion routes. Dosing errors—deviations from these protocols—pose significant risks to patient safety and trial validity (ICH Expert Working Group, 2...

work page 2001

[2] [2]

A 3,451-dimensional feature space combin- ing traditional NLP (Term Frequency–Inverse Document Frequency (TF-IDF), character n- grams), sentence embeddings, medical pat- terns, and transformer-based scores

work page

[3] [3]

Systematicanalysisshowingsentenceembed- dings and lexical features are complementary, while transformer scores underperform

work page

[4] [4]

Optuna-optimized 5-fold ensemble achieving 0.8833 ± 0.0091 CV AUC with minimal overfit- ting (0.69% out-of-fold(OOF)-test gap)

work page

[5] [5]

Automated Detection of Dosing Errors in Clinical Trial Narratives: A Multi-Modal Feature Engineering Approach with LightGBM

Feature selection improves performance: 500- 1000 features (14-29%) achieve 0.886-0.887 AUC, outperforming the full baseline (0.879) through noise reduction. arXiv:2604.19759v1 [cs.AI] 25 Mar 2026

work page internal anchor Pith review Pith/arXiv arXiv 2026

[6] [6]

Test ROC-AUC of 0.8725 on CT-DEB with threshold-adjustable recall (26-60%) for flexi- ble deployment

work page

[7] [7]

Clinical NLP and Information Extraction Clinical NLP has evolved significantly over the past two decades

Related Work 2.1. Clinical NLP and Information Extraction Clinical NLP has evolved significantly over the past two decades. Early systems like cTAKES (Savova et al., 2010) and MetaMap (Aronson and Lang,

work page 2010

[8] [8]

provided rule-based approaches for med- ical concept extraction and normalization. The i2b2 shared tasks (Uzuner et al., 2011) established benchmarks for concept extraction, assertion de- tection, and relation extraction from clinical texts, demonstrating that machine learning approaches could achieve substantial performance gains. More recent work has fo...

work page 2011

[9] [9]

Dataset and Task 3.1. CT-DEB Benchmark We utilize the CT-DEB (Clinical Trial Dosing Error Benchmark) dataset (Hêche et al., 2026), specif- ically designed for evaluating automated dosing error detection systems. The dataset comprises clinical trial narratives describing medication ad- ministration across various therapeutic areas, pro- tocols, and clinica...

work page 2026

[10] [10]

adverse event

Methodology 4.1. Data Preparation and Feature Engineering Pipeline Our feature engineering pipeline transforms raw clinical narratives into a 3,451-dimensional feature vector combining multiple representation types. Ta- ble 2 provides an overview of extracted features. Clinical trial registry data exhibits inherent spar- sity across structured fields. To ...

work page 1988

[11] [11]

Train LightGBM on the remaining 4 folds (80% of data) using weighted binary cross-entropy: L=− 1 N NX i=1 wi [yi log(ˆyi) + (1−y i) log(1−ˆyi)] (2) where wi = 20.87forpositiveexamples, wi = 1 for negative

work page

[12] [12]

Generate OOF predictions on foldk (the held- out 20%)

work page

[13] [13]

Ensemble Prediction- For test inference, all 5 fold models generate predictions, which are aver- aged: ˆyensemble(x) = 1 5 5X k=1 ˆyk(x)(3) ct-dosing-errors-benchmark

Save the trained model for later ensembling Out-of-Fold Validation- Concatenating predic- tions from all 5 folds yields complete OOF pre- dictions across the training set, providing an un- biased performance estimate: 0.8833 ± 0.0091 ROC-AUC. Ensemble Prediction- For test inference, all 5 fold models generate predictions, which are aver- aged: ˆyensemble(...

work page

[14] [14]

reduced”, “discontinued

Results 5.1. Overall Performance We employ 5-fold stratified cross-validation with ensemble averaging. Table 5 presents test set per- formance. The ROC-AUC of 0.8725 indicates excellent dis- criminativeability. Theensembledemonstratessta- ble performance with mean cross-validation AUC of 0.8833 ± 0.0091 across folds, and OOF AUC of 0.8794, showing minimal...

work page

[15] [15]

This variation reflects different train- ing/validation partitions while maintaining robust performance

when using early stopping with 200-iteration patience. This variation reflects different train- ing/validation partitions while maintaining robust performance. Cross-validation stability: Mean fold AUC of 0.883 ± 0.009 demonstrates low variance across datasplits,withindividualfoldperformanceranging from0.869to0.894(2.5%range). Thisconsistency indicates ro...

work page

[16] [16]

Most predictive value is concentrated in a small-to-moderate subset of features (500-1000), with the remaining features introducing more noise than signal

Discussion The findings indicate that accurate detection of dos- ingdeviationsinclinicalnarrativesreliesoncombin- ing sparse lexical features with contextual embed- dings,aseachcapturesdistinctbutcomplementary signals. Most predictive value is concentrated in a small-to-moderate subset of features (500-1000), with the remaining features introducing more n...

work page

[17] [17]

Conclusion We present an automated system for detecting dosing errors in clinical trial narratives, achieving 0.8725 test ROC-AUC through 5-fold ensemble learning with comprehensive multi-modal feature engineering. Our approach combines 3,451 fea- tures spanning traditional NLP (TF-IDF, character n-grams),densesemanticembeddings(all-MiniLM- L6-v2), handcr...

work page

[18] [18]

References Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. 2019. Op- tuna: A next-generation hyperparameter opti- mization framework. InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 2623–2631. Emily Alsentzer, John R. Murphy, Willie Boag, Wei-Hung Weng, Di Jin, Trist...

work page 2019

[19] [19]

clinical practice guidelines

Domain-specific language model pretrain- ing for biomedical natural language processing. ACM Transactions on Computing for Healthcare, 3(1):1–23. Rave Harpaz, William DuMouchel, Nigam H. Shah, David Madigan, Patrick Ryan, and Carol Fried- man.2012. Noveldata-miningmethodologiesfor adverse drug event discovery and analysis.Clin- ical Pharmacology & Therape...

work page arXiv 2012

[20] [20]

Özlem Uzuner, Brett R

Enhancing clinical concept extraction with contextual embeddings.Journal of the American Medical Informatics Association, 26(11):1297– 1304. Özlem Uzuner, Brett R. South, Shuying Shen, and Scott L. DuVall. 2011. 2010 i2b2/va challenge on concepts, assertions, and relations in clinical text.JournaloftheAmericanMedicalInformatics Association, 18(5):552–556....

work page 2011