Optimizing Medical Scheduling with SLMs

From Generalized AI to Hyper-Local Schedule Optimization

For local physician offices, scheduling is not a simple administrative task. Scheduling is comprised of multiple complex, high-stakes optimization problems.

Traditional, generalized AI solutions fail here because they treat a patient's "chief complaint" and a physician's "specialization" as standard text, leading to suboptimal booking, long wait times, and high no-show rates (all direct drains on EBIT).

Reason's Vertical AI platform eliminates this inefficiency by applying a focused, hyper-technical approach to a highly granular business process.

The Hyper-Focus on Encoder-Decoder Hyperparameters

Our solution utilizes a specialized Encoder-Decoder Transformer architecture optimized specifically for sequencing and constraint satisfaction—the core mechanics of scheduling. We achieve peak performance by tuning the model's hyperparameters to fit the unique rhythm of a medical office:

• Attention Head Tuning
  
  We use fewer, deeper attention heads in the Encoder to prioritize the capture of complex, conditional relationships (e.g., "This specific procedure must be followed by a 45-minute cleaning, but only requires 15 minutes of physician time").

• Decoder Constraint Layer
  
  We embed hard constraints directly into the Decoder layer's loss function. These constraints enforce non-negotiable rules such as mandated lunch breaks, regulatory limits on consecutive shift lengths, and hardware availability (e.g., "Clinic Room 3 has the specialized imaging unit, so only book a specific CPT code there").

• Sequential Prediction Optimization
  
  We fine-tune the decoding temperature to favor predictive stability over creativity, ensuring the resulting schedule sequence is not merely efficient but also highly predictable and actionable by the front-office staff.

The Power of Custom Vocabulary in Healthcare

Generalized LLMs waste resources attempting to understand the entire world. Our models utilize a custom word embedding vocabulary hyper-focused on the precise language of local medical practice.

• Clinical Taxonomy Embeddings
  
  We embed the specific ICD-10 (diagnosis) and CPT (procedure) codes unique to the office's common practice area (e.g., orthopedic surgery vs. family practice). This enables the model to instantly map a patient's natural language chief complaint ("My knee hurts when I walk") to the specific time duration and resource requirements associated with the most likely necessary CPT code.

• Resource and Staff Acronyms
  
  We incorporate the office’s internal jargon, staff names, and equipment IDs into the vocabulary. This allows the system to seamlessly translate an optimal schedule sequence into an immediately executable action plan for the front desk staff, reducing friction and training time.

The Result: Measurable Impact on EBIT

By applying this level of technical specificity, Reason's SLM platform delivers schedules that are not just full, but optimally sequenced:

• Maximized Physician Time
  
  Reduced dead time, allowing for more billable patient slots per day.

• Lower No-Show Rates
  
  Precise scheduling reduces patient frustration and wait times.

• Reduced Overtime
  
  Smoother patient flow decreases the administrative cost associated with staff overtime.