AI Case Study
Sheba Medical Center ensures accuracy of prescriptions and eliminates human error using machine learning
Medication errors are one of the biggest concerns in healthcare. Medaware's AI platform uses patient history, lab results and general prescription data to determine whether the prescription mentions the right medicine, dosage and if it is for the right patient.
Healthcare Providers And Services
" Medaware's system analysed prescriptions that passed the filtering of existing rule-based systems from 3 Israel based hospitals for prescription errors.
Retrospective analysis of the electronic medical records was performed using MedAware’s proprietary algorithms. To simulate the real-time performance and accuracy of MedAware’s system, all events (prescriptions, blood tests, admissions/discharges, etc.) were fed into the system in a temporal order, and the prescriptions were analyzed according to the accumulated data available at the time of prescription.
The types of alerts detected by MedAware’s software were all beyond the standard drug interaction, dosage or allergy related errors. They included the following errors:
Drug mix-up – prescribing the wrong drug
Patient mix-up – assigning a drug to the wrong patient
Physician unawareness of clinical data – prescribing a drug contraindicated for a patient’s status
Outliers in monitored drugs – failure to discontinue/change dose of a drug on time.
Alerts were generated in more than 1% of outpatients and in more than 3% of inpatients"
"In the US, medication errors harm at least 1.5 million people every year and cause the annual premature death of more than 220,000 patients. Adverse drug events are among the most common medical errors. Out of the 4 billion medical prescriptions that are written up annually in the US, 8 million contain life threatening errors."
"Alerts were generated in more than 1% of outpatients and in more than 3% of inpatients. In the inpatient setting, 40% of the errors were correctly identified by the nurses, thus regarded as “near-misses.” The alert burden was 1/200 inpatient prescriptions and 1/1000 outpatient prescriptions.
Patients for whom alerts were generated had a significantly longer hospital length of stay (additional 2.4 to 4 days per admission) and more hospital admissions (additional 0.6 to 1.3 annual admissions) compared to patients with no alerts."
"MedAware is differentiating itself by having alerts with high specificity and accuracy, preventing “alert fatigue”. MedAware only flags about 0.2 to 0.5% of all prescriptions, with about 75-80% being true positives, and only 35% are false negatives."
"MedAware’s engine builds a mathematical model which represents real-world treatment patterns. A prescription largely deviating from the standard treatment spectrum is likely to be erroneous"
"The inpatient cohort consisted of 23,092 patients with 33,342 hospitalizations and 476,155 prescriptions, which were admitted within a 6-8 month period. The outpatient cohort consisted of 409,546 patients with 43,647,747 prescriptions and with documented medical history of at least 5 years.
All events (prescriptions, blood tests, admissions/discharges, etc.) were fed into the system in a temporal order, and the prescriptions were analyzed according to the accumulated data available at the time of prescription"