shrirajh — publications

This page is a static rendering of pubs.json for machines and crawlers. Back to shrirajh.me

Intravenous immunoglobulin weaning evaluation with zero‐shot large language model classification

Transfusion Medicine (2026)

DOI: 10.1111/tme.70020

We applied a zero-shot LLM to classify neurology outpatients' suitability for IVIg weaning from clinic notes, achieving 78.6% classification accuracy and identifying a conservatively estimated $84,702 in potential annual cost savings from IVIg de-escalation. I helped engineer this into production. I maintain the extraction pipeline.

3567 Current specialty training selection criteria reduce medical students’ interest in neurology as a career pathway

BMJ Neurology (2025)

DOI: 10.1136/bmjno-2025-anzan.117

We surveyed medical students and found that exposure to specialty selection criteria significantly reduced interest across multiple specialties, with neurology, ICU, and ophthalmology most affected. Research requirements and competitiveness were the key deterrents. I built the analysis and visualisation tooling.

Artificial Intelligence Deployment of Conversational Support (AI-DOCS): A patient acceptability study

(2025)

DOI: 10.1101/2025.07.19.25331814

We sat an LLM alongside physician trainees in the RACP clinical exam. It scored comparably on empathy, politeness, and comprehensiveness. Notably, patients disclosed sensitive clinical information to the AI that they had withheld from human examiners. I helped engineer this into production. I maintain the conversational AI platform.

Automating cancer registries: Pearls and pitfalls

Health Information Management Journal (2025)

DOI: 10.1177/18333583251377892

We explored automating the Australian Brain Cancer Registry using AI-driven data extraction, identifying both the potential efficiency gains and practical challenges of deploying LLMs for registry data collection. I helped engineer this into production. I maintain the registry platform.

Clindamycin intravenous to oral switch automatic notifications: a pilot study

Internal Medicine Journal (2025)

DOI: 10.1111/imj.70172

We built an automated notification system to prompt clinicians when IV clindamycin could be switched to oral, and found it significantly reduced the IV-to-oral ratio and daily IV doses. I helped engineer this into production. I maintain the notification system.

Comparative Performance of Large Language Models vs Clinicians in Creating Cardiology Ward Round Notes: Promises, Limitations, and Implications for Clinical Practice

Heart, Lung and Circulation (2025)

DOI: 10.1016/j.hlc.2025.06.609

We compared LLM-generated cardiology ward round notes to clinician notes, finding LLMs produced more complete documentation but occasionally included commission errors and vague recommendations.

Computer-Assisted Protocol-Adherent Blood Lipid Evaluation in Vascular Outpatients (CAPABLE-Vascular)

Journal of Clinical Medicine (2025)

DOI: 10.3390/jcm14041321

We built a rule-based system to automatically order blood lipid and HbA1c tests before vascular surgery outpatient appointments, and found it increased test ordering from under 40% to 100% of eligible patients. I helped with the automation behind this.

Copying in medical documentation: developing an evidence‐based approach

Internal Medicine Journal (2025)

DOI: 10.1111/imj.16590

We quantified erroneous copy-paste in ward round notes, finding 8.3% contained copying errors concentrated in the issues list, and developed similarity metrics to automatically detect at-risk notes.

Electronic Quantification of Bowel Timing and Automated Alerts Reduce Inpatient Constipation

(2025)

DOI: 10.2139/ssrn.5698664

We quantified bowel motion timing electronically and delivered automated daily notifications to physicians when patients exceeded 72 hours without a bowel motion, reducing the rate of severe constipation (BNO ≥5 days) from 5.6% to 1.9%. I helped engineer this into production. I maintain the notification platform.

Epilepsy surgery candidate identification with artificial intelligence: An implementation study

Journal of Clinical Neuroscience (2025)

DOI: 10.1016/j.jocn.2025.111144

We implemented a machine learning algorithm to identify epilepsy surgery candidates from clinic data, finding that 53% of AI-flagged patients met criteria for surgery evaluation.

Large language models for infectious diseases require evidence generation and regulation

Internal Medicine Journal (2025)

DOI: 10.1111/imj.70072

We outlined the evidence generation, regulatory, and guideline requirements for deploying LLMs in infectious diseases, proposing retrieval-augmented generation as a strategy to minimise hallucination risk.

LLM-assisted medical documentation: efficacy, errors, and ethical considerations in ophthalmology

Eye (2025)

DOI: 10.1038/s41433-025-03767-5

We evaluated LLM-assisted autocomplete for ophthalmology clinic documentation, characterising error types and ethical considerations of integrating AI into clinical note-writing workflows. I built the clinical text editor and LLM integration behind this.

Medical registries: Factors impacting clinician engagement

Asia-Pacific Journal of Ophthalmology (2025)

DOI: 10.1016/j.apjo.2025.100146

We explored factors that impact clinician engagement with medical registries, identifying time burden, lack of workflow integration, and limited perceived value as key barriers to sustained participation.

Medication shortage behaviour change with multidisciplinary clinician-designed digital notification intervention

International Journal of Pharmacy Practice (2025)

DOI: 10.1093/ijpp/riae064

We deployed a clinician-designed digital notification system during an IV paracetamol shortage, and found it significantly reduced unnecessary intravenous paracetamol use.

PO-02-088 THE IMPACT OF POSTOPERATIVE ATRIAL FIBRILLATION AFTER CARDIAC SURGERY

Heart Rhythm (2025)

DOI: 10.1016/j.hrthm.2025.03.570

We studied 1,254 cardiac surgery patients and found that postoperative atrial fibrillation extended ICU stay by 0.6 days and post-ICU stay by 1.6 days, with doubled long-term mortality risk.

Predicting the Emergency Department Patient Journey Using a Machine Learning Approach

JMIR AI (2025)

DOI: 10.2196/67321

We developed a machine learning model to predict the emergency department patient journey, enabling early identification of patients likely to require admission or extended stays.

The Adelaide Score: prospective implementation of an artificial intelligence system to improve hospital and cost efficiency

ANZ Journal of Surgery (2025)

DOI: 10.1111/ans.19383

We prospectively implemented the Adelaide Score AI system to predict discharge readiness, finding it reduced median length of stay by 0.2 days and seven-day readmission rates, saving an estimated $9.5M annually. I maintain the inference server for this.

The influence of specialty training selection criteria on medical students’ career pathway choices in Australia

Australian Health Review (2025)

DOI: 10.1071/ah25085

We investigated how specialty training selection criteria influence medical students' career pathway choices in Australia, identifying which criteria most strongly shape early career decisions. I built the analysis and visualisation tooling.

When One Size Does not Fit All—Artificial Intelligence in Australian Rural Health

Australian Journal of Rural Health (2025)

DOI: 10.1111/ajr.70037

We examined the challenges of deploying AI in rural Australian hospitals, finding that connectivity limitations, smaller patient populations, and workforce constraints require tailored approaches beyond metropolitan AI models.

Zero-shot large language model application for surgical site infection auditing

Infection, Disease & Health (2025)

DOI: 10.1016/j.idh.2025.05.001

We applied a zero-shot LLM to identify surgical site infections from clinical notes, finding it could accurately flag cases without any task-specific training data or fine-tuning. I helped engineer this into production. I maintain the auditing pipeline.

Zero-shot LLM-based visual acuity extraction: a pilot study

BMC Ophthalmology (2025)

DOI: 10.1186/s12886-025-04193-7

We used a zero-shot LLM to extract visual acuity measurements from ophthalmology clinical notes, demonstrating feasibility of automated structured data extraction without task-specific training. I helped engineer this into production. I maintain the extraction pipeline.

A Generalizable Risk Factor: Socioeconomic Status and Multiresistant Organism Colonization

Clinical Infectious Diseases (2024)

DOI: 10.1093/cid/ciae651

We studied ~50,000 hospital admissions in South Australia and found that lower socioeconomic status was significantly associated with higher multiresistant organism colonisation prevalence, confirming generalisability of international findings.

Large language models can effectively extract stroke and reperfusion audit data from medical free-text discharge summaries

Journal of Clinical Neuroscience (2024)

DOI: 10.1016/j.jocn.2024.110847

We used a locally-deployed LLM to extract stroke audit data from free-text discharge summaries, achieving 93.8% accuracy across binary, categorical, datetime, and free-text fields.

Push or pull? Digital notification platform implementation reduces dysglycaemia

Internal Medicine Journal (2024)

DOI: 10.1111/imj.16506

We implemented a digital notification platform to push ahead-of-time glycaemic alerts to clinicians, and found it significantly reduced inpatient dysglycaemia events. I helped engineer this into production. I maintain the notification platform.

Rule‐based clinician‐developed programmes can facilitate haemodialysis clinical workflows

Internal Medicine Journal (2024)

DOI: 10.1111/imj.16565

We built a system that automatically detects dialysis patients admitted under non-nephrology teams and notifies the renal service before complications arise, achieving zero false positives and zero false negatives over six months. I helped engineer this into production. I maintain the patient detection system.

Suprapubic catheter change: Evaluating YouTube videos as a resource for teaching junior doctors

BJUI Compass (2024)

DOI: 10.1002/bco2.299

We evaluated YouTube videos as a teaching resource for suprapubic catheter changes, finding variable quality and identifying criteria for selecting appropriate educational content for junior doctors.

Translational artificial intelligence-led optimization and realization of estimated discharge with a supportive weekend interprofessional flow team (TAILORED-SWIFT)

Internal and Emergency Medicine (2024)

DOI: 10.1007/s11739-024-03689-2

We paired an AI discharge-prediction algorithm with a weekend interprofessional flow team, and found a significant increase in weekend discharges (18% vs 14%) with associated cost savings. I maintain the algorithm and inference server for this.