Summary
Drug Safety AI
A B2B medical AI drug safety system I designed from 0→1 at ASUS, now deployed in partner hospitals improving prescription accuracy by 28%. Core capabilities: real-time drug interaction analysis, patient allergy checking, dosage verification, and contextual AI recommendations.
Skills
User Research
UIUX Design
Prototyping
Roles
Product Designer
Collaborators
Product Manager
Design Manager
Design Mentor
Engineer
Duration
6 months
Context
Medication errors remain a critical patient safety challenge
Nearly 1 in 4 hospitalized patients experiences adverse events, with medication-related incidents accounting for over 40% of these cases.
News screenshots about the dilemma
Problem
Alert fatigue meets fragmented information
Adverse drug events (ADEs) represent one of the most critical patient safety challenges in hospitals today.
2M
Stays Were Affected
ADE in inpatient settings:
1 in 3
Hospital Adverse Events
4.6
Days
Extened Hospital Stays
The scale extends beyond inpatient settings, with millions of clinical encounters affected annually.
3.5M
Physician Office Visits
ADE in outpatient settings:
1M
Emergency Visits
125K
Hospital Admissions
Understanding where errors occur reveals the highest intervention opportunity. Prescription represents the most critical intervention point at 61% of all ADEs.
Prescribing Phase 61%
Administering Phase 22.5%
Dispensing Phase 15.7%
Other 0.8%
Error Types:
Duplicate medication orders
Incorrect dosage
Incorrect amount
Error Types:
Incorrect dosage delivery
Misinterpretation of labeling
Extravasation (IV issues)
Error Types:
Unclear or incorrect labeling
Incorrect amount dispensed
Wrong medication/formulation
Highest frequency intervention point (61% of all
ADEs). Represents the most impactful opportunity
for prevention as errors caught here prevent
downstream harm.
ADE Reasons
User interview
Understanding clinician workflow through interviews and analysis
In the flow of chatting with AI, sharing sensitive details feels natural and harmless.
We conducted in-depth interviews with healthcare providers across hospital settings to understand prescription decision-making challenges and analyzed existing workflow documentation.
Despite this clear opportunity, current systems create four critical barriers that prevent effective intervention:
Fragmented Information Systems
Patient data spreads across HIS, medical order systems, and nursing records with no unified access during prescribing
Alert Fatigue
Current systems generate excessive non-specific alerts that clinicians learn to ignore, missing truly critical interactions
Poor Timing & Context
Warnings appear without patient-specific context or actionable recommendations, requiring manual cross-referencing
Workflow Disruption
Alert systems interrupt prescribing flow without integrating into existing clinical decision-making processes
Design Principles
Alert systems that inform, not interrupt
Key Functions
Visual Relationship Discovery for Secure Enterprise Research
Based on the design principles, we came up with four key functions.
Multi-angle Drug Search
Search by brand names, generic names, symptoms, or even partial spellings.
Risk-based Alert Hierarchy
Critical alerts appear prominently with clear visual distinction; lower-priority information accessible but not disruptive.
Automated Risk Detection
System analyzes patient history, current medications, allergies, and chronic conditions from multiple data sources to identify potential interactions.
Actionable Recommendations
Instead of just flagging problems, system suggests safer alternatives based on patient profile and clinical evidence.
Ideation & Concept Development
Balancing alert urgency with workflow integration
How could we make high-risk interactions impossible to miss while keeping information accessible but non-disruptive?
Wireframe sketches
I explored multiple approaches to information hierarchy and alert presentation...from prominent modal pop-ups to subtle sidebar notifications, testing various combinations of visual weight, positioning, and interaction patterns. Early concepts included separate alert panels, embedded warnings within search results, and contextual tooltips that appeared during drug selection.
The breakthrough came from progressive disclosure within the prescription flow itself. Instead of interrupting with alerts, the system surfaces risk levels directly in the drug search results, using visual indicators (color coding and risk badges) to communicate severity at a glance. Detailed patient context and alternative recommendations expand inline only when needed, allowing quick decisions for routine prescriptions while providing deep analysis for complex cases.
Solution
Intelligent Drug Search
Flexible search handles brand names, generics, symptoms, and partial matches. "Fluen" returns Fluenz, Tamiflu, and related options with clear medication relationships.
Patient-Specific Risk Analysis
Users System automatically cross-references prescription against patient's medication history, allergies, chronic conditions, and recent lab results to identify contextual risks.choose any combination of internal systems and external platforms in a single search, eliminating the need to query multiple tools separately.
Reflection
Impact and learnings
System Performance:
87% detection rate of potential adverse drug interactions in high-risk cases during clinical trials.
Reduced medication errors in participating hospital units.
Improved prescription confidence among clinicians using the system.
What worked and why:
The 61% insight changed everything Early research revealed that most adverse drug events originate during prescription, not at other stages. This shifted our entire approach from warning systems to decision support. Instead of alerting clinicians after they'd made a choice, we could help them make better choices in the first place.
Context matters more than comprehensiveness We initially tried to show all possible drug interactions. Too much information. Clinicians don't need every theoretical risk listed, they need to know which ones actually matter for their patient right now. Patient-specific factors like kidney function or concurrent medications made the difference between "another alert to ignore" and "information I actually need."
Alert hierarchy had to feel instant Color coding and visual weight weren't just aesthetic choices. They had to communicate severity without requiring conscious thought. Red means stop and look. Orange means be aware. Green means you're fine. No reading required.
Integration was harder than expected Embedding into HIS seemed straightforward on paper. In practice, every hospital's HIS configuration was slightly different. We learned to design for flexibility rather than assuming a standard setup. The core principle stayed the same though: meet clinicians where they already work, don't make them come to us.
