Jiayu Lu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Research Interests
Jiayu Lu has a total of 9 co-authors. Here are the top co-authors by number of shared publications:
- Andrew R. Zullo
Department of Health Services, Policy, and Practice, Brown University School of Public Health, Providence, RI
2 shared publications - Diliang Chen
Department of Electrical and Computer Engineering, University of New Hampshire, Durham, NH, USA
2 shared publications - Ferdinand Delgado
Department of Kinesiology, University of New Hampshire, Durham, NH, United States
2 shared publications - Jinying Chen
Department of Medicine/Section of Preventive Medicine and Epidemiology, Boston University Chobanian & Avedisian School of Medicine, Boston, MA, USA
2 shared publications - Kevin W. McConeghy
Transformative Health Systems Research to Improve Veteran Equity and Independence Center of Innovation (THRIVE COIN), Veterans Affairs Providence Health Care System, Providence, Rhode Island, USA; Department of Health Services, Policy & Practice, Brown University School of Public Health, Providence, Rhode Island, USA
2 shared publications - Nozhan Ghoreishi
Department of Electrical and Computer Engineering, University of New Hampshire, Durham, NH, United States
2 shared publications
Latest Publications
- Open Access
- 29 September 2026
- Published by JMIR Publications Inc. in JMIR Medical Informatics
- Vol. 14, e91151
Detecting Misspelled Drug Names Using Transformer-Based Language Models: Model Development and External Validation
Background: Misspellings in medication names can compromise patient safety, reduce data utility, and impede large-scale data initiatives that integrate medication information from electronic health records (EHRs). Existing methods for detecting misspelled medical terms are mostly dictionary-based and can lead to high false-positive rates when correctly spelled but previously unseen (out-of-vocabulary) terms are encountered. Objective: We aimed to develop and validate domain-specific, transformer-based language models for detecting misspelled drug names, with an emphasis on performance for unseen terms. Methods: Using RxNorm as a standardized drug vocabulary, we created an RxNorm-augmented training corpus and developed two BERT (Bidirectional Encoder Representations from Transformers)–based models—BERTDrug and CharBERTDrug—for misspelling detection. Specifically, we randomly split 69,824 RxNorm drug names into training, development, and test sets (3:1:1) and generated k misspellings per name using text-perturbation techniques (k optimized for training; fixed at 1 for development and test sets). The models were fine-tuned on the training and development sets and evaluated using the RxNorm test set and 3586 drug names from the Long-Term Care Data Cooperative (LTCDC) database (external validation). The RxNorm test set and out-of-vocabulary LTCDC dataset (1922 terms), neither overlapping with the RxNorm training data, were used to evaluate performance on unseen terms. SpellChecker served as a dictionary-based baseline, while fastTextML and BioWordVecML, which used different subword embeddings as inputs for machine learning, served as additional baselines. Additionally, we compared model performance with GPT-4o, a generative large language model (LLM), using 2200 randomly sampled test terms. Results: On the RxNorm test set, BERTDrug and CharBERTDrug outperformed the baseline models across most metrics. BERTDrug achieved the best overall performance (F1-score=0.859; area under the receiver operating characteristic curve [ROC-AUC]=0.947), followed by CharBERTDrug (F1-score=0.833; ROC-AUC=0.906). Both models also outperformed the baseline models on the out-of-vocabulary LTCDC dataset across most metrics, with CharBERTDrug performing best (F1-score=0.696; ROC-AUC=0.788), followed by BERTDrug (F1-score=0.669; ROC-AUC=0.786). In the secondary analysis, both models exceeded GPT-4o on most metrics (except Recall) for 2000 RxNorm terms. BERTDrug performed best (ROC-AUC=0.951; F1-score=0.855), followed by CharBERTDrug (ROC-AUC=0.911; F1-score=0.831) and GPT-4o (ROC-AUC=0.856; F1-score=0.721). In contrast, among 200 randomly selected LTCDC terms, GPT-4o performed best on most metrics except precision and specificity. Conclusions: Domain-specific language models improved detection of misspellings in out-of-vocabulary drug names and outperformed baseline models in both internal and external evaluations. The comparison with a generative LLM suggests that domain shift may substantially reduce the advantages conferred by domain-specific training. With further fine-tuning on diverse data that capture the terminology, formatting conventions, and spelling patterns encountered across real-world clinical settings, these models could be adapted for use in other clinical databases and EHR systems to improve medication data quality for research and to support future safety-focused applications.
- 29 September 2026
- Published by JMIR Publications Inc. in JMIR Medical Informatics
- Vol. 14, e91151
Open AccessDetecting Misspelled Drug Names Using Transformer-Based Language Models: Model Development and External Validation
Background: Misspellings in medication names can compromise patient safety, reduce data utility, and impede large-scale data initiatives that integrate medication information from electronic health records (EHRs). Existing methods for detecting misspelled medical terms are mostly dictionary-based and can lead to high false-positive rates when correctly spelled but previously unseen (out-of-vocabulary) terms are encountered. Objective: We aimed to develop and validate domain-specific, transformer-based language models for detecting misspelled drug names, with an emphasis on performance for unseen terms. Methods: Using RxNorm as a standardized drug vocabulary, we created an RxNorm-augmented training corpus and developed two BERT (Bidirectional Encoder Representations from Transformers)–based models—BERTDrug and CharBERTDrug—for misspelling detection. Specifically, we randomly split 69,824 RxNorm drug names into training, development, and test sets (3:1:1) and generated k misspellings per name using text-perturbation techniques (k optimized for training; fixed at 1 for development and test sets). The models were fine-tuned on the training and development sets and evaluated using the RxNorm test set and 3586 drug names from the Long-Term Care Data Cooperative (LTCDC) database (external validation). The RxNorm test set and out-of-vocabulary LTCDC dataset (1922 terms), neither overlapping with the RxNorm training data, were used to evaluate performance on unseen terms. SpellChecker served as a dictionary-based baseline, while fastTextML and BioWordVecML, which used different subword embeddings as inputs for machine learning, served as additional baselines. Additionally, we compared model performance with GPT-4o, a generative large language model (LLM), using 2200 randomly sampled test terms. Results: On the RxNorm test set, BERTDrug and CharBERTDrug outperformed the baseline models across most metrics. BERTDrug achieved the best overall performance (F1-score=0.859; area under the receiver operating characteristic curve [ROC-AUC]=0.947), followed by CharBERTDrug (F1-score=0.833; ROC-AUC=0.906). Both models also outperformed the baseline models on the out-of-vocabulary LTCDC dataset across most metrics, with CharBERTDrug performing best (F1-score=0.696; ROC-AUC=0.788), followed by BERTDrug (F1-score=0.669; ROC-AUC=0.786). In the secondary analysis, both models exceeded GPT-4o on most metrics (except Recall) for 2000 RxNorm terms. BERTDrug performed best (ROC-AUC=0.951; F1-score=0.855), followed by CharBERTDrug (ROC-AUC=0.911; F1-score=0.831) and GPT-4o (ROC-AUC=0.856; F1-score=0.721). In contrast, among 200 randomly selected LTCDC terms, GPT-4o performed best on most metrics except precision and specificity. Conclusions: Domain-specific language models improved detection of misspellings in out-of-vocabulary drug names and outperformed baseline models in both internal and external evaluations. The comparison with a generative LLM suggests that domain shift may substantially reduce the advantages conferred by domain-specific training. With further fine-tuning on diverse data that capture the terminology, formatting conventions, and spelling patterns encountered across real-world clinical settings, these models could be adapted for use in other clinical databases and EHR systems to improve medication data quality for research and to support future safety-focused applications.
- Open Access
- 9 June 2026
- Published by JMIR Publications Inc. in JMIR mHealth and uHealth
- Vol. 14, e93877
Evaluation of the Importance of Stopping Elderly Accidents, Deaths, and Injuries (STEADI)–Based Factors in Wearable Fall Risk Assessment: Secondary Data Analysis
Background: Falls among older adults are a growing and costly public health problem that often leads to mobility decline and loss of independence. Although clinical frameworks such as the Centers for Disease Control and Prevention’s (CDC) Stopping Elderly Accidents, Deaths, and Injuries (STEADI) initiative recommend multifactor screening (gait, balance, strength, fear of falling, and fall history), most wearable fall risk assessment systems rely on a small set of risk factors (typically gait), which creates a gap between clinical practice and automated wearable assessment. Objective: This study aims to evaluate the importance of STEADI-based fall risk factors and provide design guidance for clinically compatible wearable fall risk assessment systems. Methods: We created a dataset of 24 older adults (10 low fall risk and 14 high fall risk) from a publicly available plantar pressure dataset of 48 participants by retaining only those with consistent fall risk labels based on both the Berg Balance Scale and the Timed Up and Go test. A total of 18 features were extracted to quantify gait, strength, balance, fear of falling, and fall history. Random forest (RF) models were trained with leave-one-subject-out cross-validation to assess fall risk. Importance of STEADI-based factors was assessed by two methods: (1) estimating Shapley Additive Explanations values based on a single RF model trained on all features; and (2) training 5 separate RF models, each on 1 STEADI factor category, and comparing their fall risk classification accuracies. Results: In this secondary analysis, the RF model trained on all features achieved a subject-level accuracy of 87.53% (95% CI 75%-100%). Shapley Additive Explanations analysis identified the right foot flat phase ratio (fear of falling feature) as the highest-ranked feature, followed by maximum right forefoot ground reaction force (strength feature), whereas traditional gait features did not appear in the top 10. The 5 separate RF models trained on individual STEADI-based factor categories showed a similar trend in mean participant-level accuracy: fear of falling, 87.59% (95% CI 75%-100%); strength, 79.18% (95% CI 62.5%-95.83%); balance, 70.5% (95% CI 50%-87.5%); gait 70.81% (95% CI 54.17%-87.5%); and fall history 62.37% (95% CI 50%-75.1%). However, paired comparisons did not show statistically significant differences in accuracy between the gait model and the models trained on other factors. Conclusions: These preliminary results show that commonly overlooked nongait factors are potentially as informative as gait, although clear superiority was not demonstrated in this dataset. The novel foot flat phase ratio ranked higher than all other evaluated features, which showed the value of domain knowledge–informed feature engineering. These preliminary findings indicate that nongait STEADI factors merit consideration in the design of wearable fall risk assessment systems.
- 9 June 2026
- Published by JMIR Publications Inc. in JMIR mHealth and uHealth
- Vol. 14, e93877
Open AccessEvaluation of the Importance of Stopping Elderly Accidents, Deaths, and Injuries (STEADI)–Based Factors in Wearable Fall Risk Assessment: Secondary Data Analysis
Background: Falls among older adults are a growing and costly public health problem that often leads to mobility decline and loss of independence. Although clinical frameworks such as the Centers for Disease Control and Prevention’s (CDC) Stopping Elderly Accidents, Deaths, and Injuries (STEADI) initiative recommend multifactor screening (gait, balance, strength, fear of falling, and fall history), most wearable fall risk assessment systems rely on a small set of risk factors (typically gait), which creates a gap between clinical practice and automated wearable assessment. Objective: This study aims to evaluate the importance of STEADI-based fall risk factors and provide design guidance for clinically compatible wearable fall risk assessment systems. Methods: We created a dataset of 24 older adults (10 low fall risk and 14 high fall risk) from a publicly available plantar pressure dataset of 48 participants by retaining only those with consistent fall risk labels based on both the Berg Balance Scale and the Timed Up and Go test. A total of 18 features were extracted to quantify gait, strength, balance, fear of falling, and fall history. Random forest (RF) models were trained with leave-one-subject-out cross-validation to assess fall risk. Importance of STEADI-based factors was assessed by two methods: (1) estimating Shapley Additive Explanations values based on a single RF model trained on all features; and (2) training 5 separate RF models, each on 1 STEADI factor category, and comparing their fall risk classification accuracies. Results: In this secondary analysis, the RF model trained on all features achieved a subject-level accuracy of 87.53% (95% CI 75%-100%). Shapley Additive Explanations analysis identified the right foot flat phase ratio (fear of falling feature) as the highest-ranked feature, followed by maximum right forefoot ground reaction force (strength feature), whereas traditional gait features did not appear in the top 10. The 5 separate RF models trained on individual STEADI-based factor categories showed a similar trend in mean participant-level accuracy: fear of falling, 87.59% (95% CI 75%-100%); strength, 79.18% (95% CI 62.5%-95.83%); balance, 70.5% (95% CI 50%-87.5%); gait 70.81% (95% CI 54.17%-87.5%); and fall history 62.37% (95% CI 50%-75.1%). However, paired comparisons did not show statistically significant differences in accuracy between the gait model and the models trained on other factors. Conclusions: These preliminary results show that commonly overlooked nongait factors are potentially as informative as gait, although clear superiority was not demonstrated in this dataset. The novel foot flat phase ratio ranked higher than all other evaluated features, which showed the value of domain knowledge–informed feature engineering. These preliminary findings indicate that nongait STEADI factors merit consideration in the design of wearable fall risk assessment systems.
- Open Access
- 23 February 2026
- Published by JMIR Publications Inc. in JMIR Preprints
Evaluation of the Importance of Stopping Elderly Accidents, Deaths, and Injuries (STEADI)–Based Factors in Wearable Fall Risk Assessment: Secondary Data Analysis (Preprint)
BACKGROUND Falls among older adults are a growing and costly public health problem that often leads to mobility decline and loss of independence. Although clinical frameworks such as the Centers for Disease Control and Prevention’s (CDC) Stopping Elderly Accidents, Deaths, and Injuries (STEADI) initiative recommend multifactor screening (gait, balance, strength, fear of falling, and fall history), most wearable fall risk assessment systems rely on a small set of risk factors (typically gait), which creates a gap between clinical practice and automated wearable assessment. OBJECTIVE This study aims to evaluate the importance of STEADI-based fall risk factors and provide design guidance for clinically compatible wearable fall risk assessment systems. METHODS We created a dataset of 24 older adults (10 low fall risk and 14 high fall risk) from a publicly available plantar pressure dataset of 48 participants by retaining only those with consistent fall risk labels based on both the Berg Balance Scale and the Timed Up and Go test. A total of 18 features were extracted to quantify gait, strength, balance, fear of falling, and fall history. Random forest (RF) models were trained with leave-one-subject-out cross-validation to assess fall risk. Importance of STEADI-based factors was assessed by two methods: (1) estimating Shapley Additive Explanations values based on a single RF model trained on all features; and (2) training 5 separate RF models, each on 1 STEADI factor category, and comparing their fall risk classification accuracies. RESULTS In this secondary analysis, the RF model trained on all features achieved a subject-level accuracy of 87.53% (95% CI 75%-100%). Shapley Additive Explanations analysis identified the right foot flat phase ratio (fear of falling feature) as the highest-ranked feature, followed by maximum right forefoot ground reaction force (strength feature), whereas traditional gait features did not appear in the top 10. The 5 separate RF models trained on individual STEADI-based factor categories showed a similar trend in mean participant-level accuracy: fear of falling, 87.59% (95% CI 75%-100%); strength, 79.18% (95% CI 62.5%-95.83%); balance, 70.5% (95% CI 50%-87.5%); gait 70.81% (95% CI 54.17%-87.5%); and fall history 62.37% (95% CI 50%-75.1%). However, paired comparisons did not show statistically significant differences in accuracy between the gait model and the models trained on other factors. CONCLUSIONS These preliminary results show that commonly overlooked nongait factors are potentially as informative as gait, although clear superiority was not demonstrated in this dataset. The novel foot flat phase ratio ranked higher than all other evaluated features, which showed the value of domain knowledge–informed feature engineering. These preliminary findings indicate that nongait STEADI factors merit consideration in the design of wearable fall risk assessment systems.
- 23 February 2026
- Published by JMIR Publications Inc. in JMIR Preprints
Open AccessEvaluation of the Importance of Stopping Elderly Accidents, Deaths, and Injuries (STEADI)–Based Factors in Wearable Fall Risk Assessment: Secondary Data Analysis (Preprint)
BACKGROUND Falls among older adults are a growing and costly public health problem that often leads to mobility decline and loss of independence. Although clinical frameworks such as the Centers for Disease Control and Prevention’s (CDC) Stopping Elderly Accidents, Deaths, and Injuries (STEADI) initiative recommend multifactor screening (gait, balance, strength, fear of falling, and fall history), most wearable fall risk assessment systems rely on a small set of risk factors (typically gait), which creates a gap between clinical practice and automated wearable assessment. OBJECTIVE This study aims to evaluate the importance of STEADI-based fall risk factors and provide design guidance for clinically compatible wearable fall risk assessment systems. METHODS We created a dataset of 24 older adults (10 low fall risk and 14 high fall risk) from a publicly available plantar pressure dataset of 48 participants by retaining only those with consistent fall risk labels based on both the Berg Balance Scale and the Timed Up and Go test. A total of 18 features were extracted to quantify gait, strength, balance, fear of falling, and fall history. Random forest (RF) models were trained with leave-one-subject-out cross-validation to assess fall risk. Importance of STEADI-based factors was assessed by two methods: (1) estimating Shapley Additive Explanations values based on a single RF model trained on all features; and (2) training 5 separate RF models, each on 1 STEADI factor category, and comparing their fall risk classification accuracies. RESULTS In this secondary analysis, the RF model trained on all features achieved a subject-level accuracy of 87.53% (95% CI 75%-100%). Shapley Additive Explanations analysis identified the right foot flat phase ratio (fear of falling feature) as the highest-ranked feature, followed by maximum right forefoot ground reaction force (strength feature), whereas traditional gait features did not appear in the top 10. The 5 separate RF models trained on individual STEADI-based factor categories showed a similar trend in mean participant-level accuracy: fear of falling, 87.59% (95% CI 75%-100%); strength, 79.18% (95% CI 62.5%-95.83%); balance, 70.5% (95% CI 50%-87.5%); gait 70.81% (95% CI 54.17%-87.5%); and fall history 62.37% (95% CI 50%-75.1%). However, paired comparisons did not show statistically significant differences in accuracy between the gait model and the models trained on other factors. CONCLUSIONS These preliminary results show that commonly overlooked nongait factors are potentially as informative as gait, although clear superiority was not demonstrated in this dataset. The novel foot flat phase ratio ranked higher than all other evaluated features, which showed the value of domain knowledge–informed feature engineering. These preliminary findings indicate that nongait STEADI factors merit consideration in the design of wearable fall risk assessment systems.
- Open Access
- 19 January 2026
- Published by JMIR Publications Inc. in JMIR Preprints
Detecting Misspelled Drug Names Using Transformer-Based Language Models: Model Development and External Validation (Preprint)
BACKGROUND Misspellings in medication names can compromise patient safety, reduce data utility, and impede large-scale data initiatives that integrate medication information from electronic health records (EHRs). Existing methods for detecting misspelled medical terms are mostly dictionary-based and can lead to high false-positive rates when correctly spelled but previously unseen (out-of-vocabulary) terms are encountered. OBJECTIVE We aimed to develop and validate domain-specific, transformer-based language models for detecting misspelled drug names, with an emphasis on performance for unseen terms. METHODS Using RxNorm as a standardized drug vocabulary, we created an RxNorm-augmented training corpus and developed two BERT (Bidirectional Encoder Representations from Transformers)–based models—BERTDrug and CharBERTDrug—for misspelling detection. Specifically, we randomly split 69,824 RxNorm drug names into training, development, and test sets (3:1:1) and generated k misspellings per name using text-perturbation techniques (k optimized for training; fixed at 1 for development and test sets). The models were fine-tuned on the training and development sets and evaluated using the RxNorm test set and 3586 drug names from the Long-Term Care Data Cooperative (LTCDC) database (external validation). The RxNorm test set and out-of-vocabulary LTCDC dataset (1922 terms), neither overlapping with the RxNorm training data, were used to evaluate performance on unseen terms. SpellChecker served as a dictionary-based baseline, while fastTextML and BioWordVecML, which used different subword embeddings as inputs for machine learning, served as additional baselines. Additionally, we compared model performance with GPT-4o, a generative large language model (LLM), using 2200 randomly sampled test terms. RESULTS On the RxNorm test set, BERTDrug and CharBERTDrug outperformed the baseline models across most metrics. BERTDrug achieved the best overall performance (F1-score=0.859; area under the receiver operating characteristic curve [ROC-AUC]=0.947), followed by CharBERTDrug (F1-score=0.833; ROC-AUC=0.906). Both models also outperformed the baseline models on the out-of-vocabulary LTCDC dataset across most metrics, with CharBERTDrug performing best (F1-score=0.696; ROC-AUC=0.788), followed by BERTDrug (F1-score=0.669; ROC-AUC=0.786). In the secondary analysis, both models exceeded GPT-4o on most metrics (except Recall) for 2000 RxNorm terms. BERTDrug performed best (ROC-AUC=0.951; F1-score=0.855), followed by CharBERTDrug (ROC-AUC=0.911; F1-score=0.831) and GPT-4o (ROC-AUC=0.856; F1-score=0.721). In contrast, among 200 randomly selected LTCDC terms, GPT-4o performed best on most metrics except precision and specificity. CONCLUSIONS Domain-specific language models improved detection of misspellings in out-of-vocabulary drug names and outperformed baseline models in both internal and external evaluations. The comparison with a generative LLM suggests that domain shift may substantially reduce the advantages conferred by domain-specific training. With further fine-tuning on diverse data that capture the terminology, formatting conventions, and spelling patterns encountered across real-world clinical settings, these models could be adapted for use in other clinical databases and EHR systems to improve medication data quality for research and to support future safety-focused applications.
- 19 January 2026
- Published by JMIR Publications Inc. in JMIR Preprints
Open AccessDetecting Misspelled Drug Names Using Transformer-Based Language Models: Model Development and External Validation (Preprint)
BACKGROUND Misspellings in medication names can compromise patient safety, reduce data utility, and impede large-scale data initiatives that integrate medication information from electronic health records (EHRs). Existing methods for detecting misspelled medical terms are mostly dictionary-based and can lead to high false-positive rates when correctly spelled but previously unseen (out-of-vocabulary) terms are encountered. OBJECTIVE We aimed to develop and validate domain-specific, transformer-based language models for detecting misspelled drug names, with an emphasis on performance for unseen terms. METHODS Using RxNorm as a standardized drug vocabulary, we created an RxNorm-augmented training corpus and developed two BERT (Bidirectional Encoder Representations from Transformers)–based models—BERTDrug and CharBERTDrug—for misspelling detection. Specifically, we randomly split 69,824 RxNorm drug names into training, development, and test sets (3:1:1) and generated k misspellings per name using text-perturbation techniques (k optimized for training; fixed at 1 for development and test sets). The models were fine-tuned on the training and development sets and evaluated using the RxNorm test set and 3586 drug names from the Long-Term Care Data Cooperative (LTCDC) database (external validation). The RxNorm test set and out-of-vocabulary LTCDC dataset (1922 terms), neither overlapping with the RxNorm training data, were used to evaluate performance on unseen terms. SpellChecker served as a dictionary-based baseline, while fastTextML and BioWordVecML, which used different subword embeddings as inputs for machine learning, served as additional baselines. Additionally, we compared model performance with GPT-4o, a generative large language model (LLM), using 2200 randomly sampled test terms. RESULTS On the RxNorm test set, BERTDrug and CharBERTDrug outperformed the baseline models across most metrics. BERTDrug achieved the best overall performance (F1-score=0.859; area under the receiver operating characteristic curve [ROC-AUC]=0.947), followed by CharBERTDrug (F1-score=0.833; ROC-AUC=0.906). Both models also outperformed the baseline models on the out-of-vocabulary LTCDC dataset across most metrics, with CharBERTDrug performing best (F1-score=0.696; ROC-AUC=0.788), followed by BERTDrug (F1-score=0.669; ROC-AUC=0.786). In the secondary analysis, both models exceeded GPT-4o on most metrics (except Recall) for 2000 RxNorm terms. BERTDrug performed best (ROC-AUC=0.951; F1-score=0.855), followed by CharBERTDrug (ROC-AUC=0.911; F1-score=0.831) and GPT-4o (ROC-AUC=0.856; F1-score=0.721). In contrast, among 200 randomly selected LTCDC terms, GPT-4o performed best on most metrics except precision and specificity. CONCLUSIONS Domain-specific language models improved detection of misspellings in out-of-vocabulary drug names and outperformed baseline models in both internal and external evaluations. The comparison with a generative LLM suggests that domain shift may substantially reduce the advantages conferred by domain-specific training. With further fine-tuning on diverse data that capture the terminology, formatting conventions, and spelling patterns encountered across real-world clinical settings, these models could be adapted for use in other clinical databases and EHR systems to improve medication data quality for research and to support future safety-focused applications.