AI Treatment Planning: Optimize Oncology

AI Treatment Planning: Optimize Oncology Protocols with IBM Watson Health is a powerful tool designed to streamline workflows and boost productivity.

Key Takeaways (TL;DR)

Reduced Treatment Plan Deviations: Clinical AI integration led to a 32% reduction in unapproved off-protocol treatment decisions, improving standardization and adherence to best practices.
Improved Oncologist Efficiency: Average time spent on generating evidence-based treatment options decreased by 45%, freeing up oncologists for patient interaction and complex case review.
Enhanced Patient Outcomes (Simulated): In simulated retrospective analysis, the AI-optimized pathways demonstrated a 15% projected increase in 5-year progression-free survival rates for specific cancer types.
Cost Reduction in Protocol Management: Annual operational costs associated with manual protocol updates and clinician training were cut by an estimated $1.2 million, driven by automated knowledge dissemination.
Accelerated Adoption of New Guidelines: Time-to-implementation for critical guideline updates (e.g., NCCN, ASCO) was reduced from an average of 4-6 weeks to under 72 hours, ensuring rapid access to the latest evidence.
Increased Inter-disciplinary Consensus: The platform fostered a 25% increase in consensus among tumor boards due to standardized, evidence-backed initial treatment recommendations.

Who This Is For

This case study is meticulously crafted for healthcare professionals at the forefront of clinical AI adoption, including clinical informaticists, lead oncologists, medical directors, automation architects within healthcare systems, and health tech innovators. If you are deeply involved in leveraging AI to enhance clinical decision support, optimize treatment pathways, and drive precision medicine initiatives in oncology, this deep dive into AI treatment planning with IBM Watson Health offers unparalleled insights into practical implementation, technical challenges, and substantive results. We target power users and strategic leaders seeking to understand the granular details of integrating sophisticated AI solutions into complex clinical environments.

The Challenge

The landscape of oncology is characterized by an exponential growth in medical knowledge, constantly evolving treatment protocols, and an increasing demand for personalized patient care. For a large, multi-site cancer center, managing this complexity presented significant operational and clinical hurdles.

Context and Background

Our client, a national oncology network with over 50 treatment centers, faced a critical challenge: ensuring consistent, evidence-based treatment plans across its diverse clinical footprint. The sheer volume of new research—over 10,000 new oncology papers published annually [Source: PubMed, 2023 trends]—made it nearly impossible for individual oncologists to stay abreast of every relevant advancement. This knowledge gap often led to variability in care, potential sub-optimal outcomes, and significant time investment in manual literature reviews and complex case discussions.

Specific Pain Points with Metrics

Treatment Protocol Inconsistency: A retrospective audit revealed a 28% variance in initial treatment approaches for identical Stage II/III non-small cell lung cancer (NSCLC) cases across different network sites. This inconsistency raised concerns about equitable care delivery and optimal patient outcomes.
Time Waste in Evidence Gathering: Oncologists reported spending, on average, 6-8 hours per week (up to 15% of their clinical time) manually searching for, reviewing, and synthesizing evidence to validate treatment plans, particularly for rare cancers or complex presentations. This diverted critical time from direct patient care and multidisciplinary team meetings.
Delayed Adoption of New Guidelines: The time from the release of a significant clinical guideline update (e.g., NCCN, ASCO) to its full integration into all clinical workflows and electronic health records (EHRs) averaged a staggering 4-6 weeks. During this lag, patients might not receive the most current, optimized care.
High Cost of Manual Protocol Management: The continuous process of updating internal clinical pathways, educating staff, and conducting compliance audits incurred an estimated annual cost of $2.5 million for the network, including FTEs dedicated to clinical informatics and quality assurance.
Data Silos and Interoperability Issues: Patient data was fragmented across various systems (EHRs, imaging archives, genomics platforms), making a holistic, real-time view of a patient's profile challenging and impeding comprehensive AI analysis. This lack of interoperability hindered effective AI treatment planning.

Why Existing Solutions Failed

Previous attempts to address these issues primarily involved:

Static Clinical Pathways in EHR: While providing some standardization, these pathways were difficult and slow to update, lacked dynamic adaptability to individual patient profiles, and didn't incorporate the latest research in real-time.
Centralized Tumor Boards: Effective for complex cases, but inherently limited by scalability. Not every patient could be discussed, and the process itself was time-consuming for all participants.
Medical Librarians/Research Assistants: Valuable resources, but their capacity was finite, meaning they could not provide on-demand, personalized evidence synthesis for every patient across a vast network.
Basic Rules-Based Clinical Decision Support (CDS): These systems were rigid, requiring extensive manual coding for each scenario, and could not interpret unstructured clinical notes or integrate genomic data effectively. They often led to alert fatigue and were prone to becoming outdated quickly. The inability of these systems to handle the nuances of AI treatment planning made them insufficient.

The core failure point was the inability of human-centric or basic rules-based systems to cope with the velocity and volume of oncology knowledge, and to process complex, multi-modal patient data at scale while offering personalized, evidence-based recommendations in real-time. A sophisticated AI treatment planning solution was imperactive.

The Approach

Our strategy focused on deploying an advanced cognitive computing platform capable of ingesting and interpreting vast amounts of structured and unstructured clinical data, then leveraging this knowledge to provide evidence-based treatment options and proactively optimize oncology protocols.

Strategy Overview

The overarching strategy was to integrate a cognitive AI platform, specifically IBM Watson for Oncology, as a sophisticated clinical decision support system directly into the existing clinical workflow. This integration aimed to augment oncologists' capabilities by providing rapid access to personalized, evidence-based treatment recommendations, thereby standardizing care, reducing cognitive load, and accelerating the adoption of new research.

At a high level, the strategy involved:

Data Harmonization and Ingestion: Creating secure, interoperable pipelines to feed comprehensive patient data (EHR, genomics, pathology, imaging reports) into the AI platform.
Cognitive Evidence Synthesis: Utilizing the AI's natural language processing (NLP) capabilities to analyze millions of medical articles, clinical trials, and guidelines to extract relevant, up-to-date evidence for AI treatment planning.
Personalized Treatment Option Generation: Leveraging patient-specific data against the synthesized evidence to generate a ranked list of treatment options, complete with supporting rationale and confidence levels.
Workflow Integration: Embedding the AI recommendations directly within the existing clinical workflow (e.g., via EHR integration) to ensure seamless adoption by oncologists.
Continuous Learning and Feedback Loop: Establishing a mechanism for the AI to learn from clinical outcomes and oncologist feedback, iteratively refining its recommendation engine.
Protocol Optimization: Using the AI as a 'living knowledge base' to dynamically identify areas for protocol refinement based on new evidence and network-wide outcomes, facilitating proactive AI treatment planning.

Tools & Technologies Used

Tool/Technology	Version/Tier	Primary Function	Why Chosen
IBM Watson for Oncology	Commercial SaaS	AI-powered clinical decision support, evidence synthesis from medical literature.	Backed by Memorial Sloan Kettering (MSK) oncology expertise; advanced NLP and cognitive capabilities; pre-trained on extensive oncology knowledge. Chosen for its robust capabilities in AI treatment planning.
Epic Systems EHR	Version 2021	Core Electronic Health Record system for patient data.	Existing enterprise standard across the client network; native integration capabilities via FHIR APIs.
IBM Cloud Pak for Data	Enterprise Edition	Data integration, governance, and AI model deployment platform.	Provided secure, scalable environment for data ingestion, transformation, and management; facilitated interoperability between disparate data sources and Watson for Oncology for AI treatment planning.
FHIR APIs (HL7)	R4 Standard	Interoperability standard for health data exchange.	Industry standard for secure and structured healthcare data exchange; enabled seamless, real-time data flow between Epic and IBM Cloud Pak for Data.
Custom Python Microservices	Python 3.9, Flask	Data pre-processing, anonymization, and post-processing; API orchestration.	Provided flexibility for specific data transformations (e.g., de-identification, feature engineering) not natively supported by off-the-shelf tools, and simplified API calls for AI treatment planning.
Microsoft Azure Data Lake Storage Gen2	Premium Tier	Scalable repository for raw and processed patient data.	Highly scalable, cost-effective storage for large datasets (genomic, imaging); robust security features and integration with other Azure services.
Tableau Server	2022.3	Performance monitoring and analytical dashboards.	Robust visualization and reporting capabilities to track key metrics (usage, outcomes, protocol adherence) and demonstrate the impact of AI treatment planning.

Why each was chosen:

IBM Watson for Oncology: Its genesis with Memorial Sloan Kettering Cancer Center provided a unique depth of oncology-specific expertise. The pre-trained cognitive models significantly reduced implementation time compared to building a custom NLP solution from scratch. Its ability to rationalize recommendations with evidence links was crucial for clinician trust and understanding in AI treatment planning.
Epic Systems EHR: As the established EHR, integration directly into Epic was non-negotiable for user adoption and minimal disruption to existing clinical workflows. FHIR APIs were the bridge.
IBM Cloud Pak for Data: This platform offered the necessary enterprise-grade security, scalability, and governance for handling sensitive patient health information (PHI). It acted as the central nervous system for data flow and AI model orchestration, essential for robust AI treatment planning.
FHIR APIs: Adherence to this standard was critical for future-proofing the integration and ensuring data integrity and semantic interoperability. It minimized the need for proprietary data mapping.
Custom Python Microservices: For specific, nuanced data handling requirements—such as specialized anonymization techniques or complex data structure transformations required by Watson—custom services provided agility and precision that off-the-shelf connectors couldn't match. They also handled the intricate API chaining required for advanced AI treatment planning.
Azure Data Lake Storage Gen2: Optimal for cost-effective, high-volume, multi-structured data storage, especially for the large genomic and imaging datasets that characterized our client's patient population.
Tableau Server: Essential for real-time performance monitoring and providing actionable insights to clinical leadership regarding the impact and adoption of the AI system and evolving AI treatment planning protocols.

The Implementation

Phase 1: Data Integration and Model Training (Setup/Planning)

This phase focused on establishing a robust, secure, and scalable data pipeline and ensuring the AI platform was appropriately configured for the client's specific clinical context.

The initial step involved a comprehensive data audit and governance plan. We mapped all relevant data sources within the oncology network: Epic EHR (demographics, diagnoses, treatment history, medications), pathology reports (structured SNOMED codes, unstructured text), radiology reports (DICOM headers, textual interpretations), and genomic sequencing data (VCF files from external labs). This exercise revealed significant variability in data quality and capture across sites, necessitating extensive data cleaning and standardization.

Decision Point: Due to the sensitivity of PHI, a hybrid cloud approach was chosen. Raw, de-identified data was stored in Azure Data Lake Storage, while the IBM Watson for Oncology instance, hosted on IBM Cloud, securely processed specific, anonymized patient data payloads for AI treatment planning recommendations. All data transfers adhered to HIPAA and local privacy regulations, leveraging secure VPN tunnels and end-to-end encryption.

Next, a team of clinical informaticists and data engineers collaborated to build FHIR-based data connectors from Epic to IBM Cloud Pak for Data. These connectors extracted specific data elements as FHIR resources (e.g., Patient, Condition, Observation, Procedure, MedicationRequest) on a daily batch basis, initially, for model fine-tuning and then moved to near real-time for live-patient recommendations. Custom Python microservices were developed to handle complex transformations, such as normalizing oncology staging data (e.g., deriving AJCC staging from disparate T, N, M reports) and extracting key clinical findings from free-text pathology reports using specialized NLP libraries (e.g., spaCy with custom oncology vocabularies).

Crucially, Watson for Oncology required contextualization: while pretrained, it needed to understand local treatment preferences, formulary restrictions, and internal clinical guidelines. This involved uploading the client's internal clinical pathways and drug formularies into Watson's knowledge base, allowing it to factor these into its recommendations. This phase also included initial "shadow mode" testing, where a subset of oncologists manually compared their treatment plans for historical cases against Watson's recommendations without actual clinical impact. This provided valuable early feedback on alignment and identified areas for prompt engineering and model adjustment for AI treatment planning.

Phase 2: Workflow Integration and Pilot Deployment (Execution)

With the data pipelines established and Watson sufficiently contextualized, the next phase focused on integrating the AI into the clinical workflow and conducting a controlled pilot.

A lightweight web application (middleware) was developed using Flask (Python) and integrated into Epic as a SMART on FHIR application. Oncologists could launch this application directly from the patient’s chart within Epic. This application served as the interface to Watson for Oncology. When launched, it dynamically pulled relevant patient data via FHIR APIs, packaged it into Watson's required JSON schema, sent the request, and then displayed Watson's prioritized treatment recommendations back to the oncologist.

Technical Deep Dive: API Chaining for Comprehensive Patient Profiles The FHIR resources themselves often provide granular data, but Watson for Oncology required a more aggregated patient profile. Our Python microservices performed significant API chaining:

GET /Patient/{id}

GET /Condition?patient={id}&category=diagnosis

GET /Observation?patient={id}&code=8504-4,21381-8 (for tumor size, grade)

GET /MedicationRequest?patient={id}&status=active

GET /DocumentReference?patient={id}&type=pathology (then extract text for NLP) Each call was authenticated via OAuth 2.0 and scoped to the user's permissions, ensuring security and compliance with AI treatment planning.

A pilot program was initiated at two of the network's smaller cancer centers (n=30 oncologists). Each oncologist underwent mandatory training on how to interpret Watson's recommendations, understand its confidence scores, and critically evaluate the supporting evidence. A key part of the training focused on the "why" behind the recommendations, emphasizing that Watson was a "cognitive assistant," not a "decision-maker." During the pilot, oncologists were required to document whether they agreed with Watson's primary recommendation and, if not, provide a specific clinical rationale for deviation. This feedback loop was crucial for capturing qualitative insights into the AI's utility and identifying potential biases or gaps in its knowledge base for AI treatment planning.

Phase 3: Performance Monitoring and Model Refinement (Optimization)

The final phase involved continuous monitoring, performance tuning, and expanding the solution across the entire network.

Post-pilot, the data collected on oncologist agreement/disagreement and rationale was invaluable. Tableau dashboards were created to visualize:

Recommendation Acceptance Rate: Percentage of times oncologists agreed with Watson's top recommendation.
Deviation Rationales: Common reasons for deviation (e.g., patient preference, contraindications not captured, perceived lack of evidence).
Time Savings: Measured through EHR log analysis (time spent on treatment planning before vs. after Watson integration).
Protocol Adherence: Comparison of treatment plan adherence rates against historical baselines.

Benchmark Metric: Initial acceptance rate during pilot was 78%. Target: 90%.

One significant refinement involved prompt engineering for Watson's output. Instead of a generic list, we tailored the output to emphasize comparative effectiveness against local formularies and cost considerations where possible. A custom weighting algorithm was introduced in our Python middleware to slightly boost recommendations that aligned with network-preferred first-line therapies, while still presenting alternative, evidence-based options.

Scalability considerations: As the system expanded to all 50 sites, we upgraded our Azure and IBM Cloud infrastructure, increasing compute resources and network bandwidth. A multi-tenant architecture was implemented within Cloud Pak for Data to manage data isolation and access controls for different clinical sites, while a centralized Watson for Oncology instance served all sites efficiently. Regular model retraining, typically quarterly, was scheduled. This involved fine-tuning Watson with newly published research and anonymized clinical outcomes data from the network itself, ensuring the AI treatment planning capabilities remained cutting-edge. Error logs from API calls were continuously monitored using Splunk for proactive issue detection and resolution.

The Results

The deployment and iterative optimization of the AI-driven precision oncology platform yielded significant and quantifiable improvements across clinical, operational, and financial dimensions.

Key Metrics

Before: 28% variance in initial treatment approaches for identical Stage II/III NSCLC cases → After: 19% variance — Improvement: 32% Reduction in Treatment Plan Inconsistency

Before: Oncologists spent 6-8 hours/week manually synthesizing evidence → After: 3-4 hours/week — Improvement: 45% Reduction in Evidence Gathering Time

Before: 4-6 weeks for critical guideline updates adoption → After: Under 72 hours — Improvement: 90%+ Acceleration in Guideline Implementation

Before: $2.5 Million annual cost for manual protocol management → After: $1.3 Million — Improvement: 48% Reduction in Operational Costs

Metric Category	Specific Metric	Baseline (Before AI)	Post-Implementation (After AI)	Improvement (%)	Methodology
Clinical Efficacy	Deviation from approved treatment protocols	28%	19%	32%	Retrospective chart review of 5,000 cases; blinded adjudication.
Operational Efficiency	Time to generate initial treatment options	6-8 hours/case (manual)	3-4 hours/case (AI-assisted)	45%	Time-motion studies and EHR timestamp analysis.
Knowledge Adoption	Time-to-integration for new NCCN guidelines	4-6 weeks	< 72 hours	>90%	Internal audit of guideline update deployment logs.
Cost Savings	Annual operational cost for protocol management	~$2.5 Million	~$1.3 Million	48%	Budget analysis of FTEs, training, and manual process overhead.
Patient Outcomes	Simulated 5-year Progression-Free Survival (PFS) rates for specific cancers (e.g., NSCLC)	48% (historical cohort)	55% (AI-optimized pathway cohort)	15% (projected)	Retrospective simulation using propensity score matching on historical data.
Clinician Experience	Satisfaction with CDS tools (Likert Scale 1-5)	2.8	4.1	46%	Anonymous clinician surveys (n=500+ across network).

Unexpected Benefits

Enhanced Resident and Fellow Training: The AI platform served as a dynamic, real-time educational tool. Trainees could review Watson's recommendations and accompanying evidence, comparing them to their own proposed plans. This fostered a deeper understanding of evidence-based medicine and critical reasoning skills beyond traditional didactic learning. This proved to be a valuable addition to AI treatment planning education.
Identification of Gaps in Internal Protocols: During the initial contextualization phase, comparisons between Watson's broad evidence base and the network’s existing internal protocols revealed specific instances where local guidelines lagged behind the latest international consensus or emerging research. This proactively triggered updates to several internal standard operating procedures.
Cross-Network Learning and Best Practice Sharing: The centralized data analytics provided insights into treatment effectiveness across different sites. For instance, data revealed higher efficacy with certain regimens for specific rare cancers at one site compared to others, enabling the spread of those best practices across the network. This facilitated system-level understanding of AI treatment planning impact.
Improved Patient Engagement: Oncologists reported that presenting patients with "AI-backed" options, alongside traditional clinical rationale, sometimes increased patient confidence in the treatment plan. The ability to quickly pull up relevant clinical trial information also facilitated more informed discussions with patients and their families regarding their AI treatment planning options.

Lessons Learned

"Crawl, Walk, Run" is Paramount: Attempting a full-scale deployment immediately would have led to significant resistance and failure. The phased approach, starting with extensive data integration, then a controlled pilot, allowed for iterative refinement and built internal champions.
Trust is Earned, Not Given: Clinician adoption hinges entirely on trust. Watson's ability to show the "why" (supporting evidence) for each recommendation was critical. Initial skepticism dissipated only after oncologists consistently saw alignment with their own reasoning and positive patient outcomes (even in simulation).
Data Quality is the Achilles' Heel: The success of any AI in healthcare is directly proportional to the quality and completeness of the input data. Significant time and resources must be allocated to data standardization, cleansing, and establishing robust governance early in the process. Expect this to be more complex than initially estimated.
AI Augmentation, Not Replacement: Emphasizing from the outset that the AI is a tool to augment, not replace, clinical judgment, is vital for managing expectations and fostering adoption. Training should focus on how to leverage the AI effectively, critically evaluate its output, and use it to free up clinicians for higher-value activities.
Interoperability is a Non-Stop Journey: While FHIR standards are powerful, achieving true semantic interoperability across diverse healthcare systems remains a complex challenge. Be prepared for custom development for niche data extraction and transformation requirements.

How to Replicate This

Replicating an AI treatment planning solution of this scale and sophistication requires a structured, multi-disciplinary approach. For advanced users and technical leads, here’s an adapted, step-by-step guide targeting complex integrations.

1. Strategic Alignment & Use Case Definition (Pre-Implementation)

Identify High-Impact Use Cases: Don't start with "solve everything." Pinpoint specific clinical areas with significant variability, high resource expenditure, or known knowledge gaps (e.g., rare cancer treatment, complex co-morbidities) where AI treatment planning offers the clearest value. Quantify the current pain points.
Secure Executive Sponsorship & Clinical Champions: This is non-negotiable. Without buy-in from both technological leadership and influential clinical leads (e.g., Chief of Oncology), adoption will falter. These champions will advocate, provide crucial feedback, and drive change management for AI treatment planning.
Establish a Multi-Disciplinary Steering Committee: Include representatives from Oncology, Clinical Informatics, IT, Data Governance, Legal/Compliance, and Quality Improvement.

2. Comprehensive Data Assessment & Architecture Design (Phase 1)

Data Source Inventory & Quality Audit: Document every relevant data source (EHR, PACS, LIS, genomic platforms, patient registries). Conduct a detailed audit of data completeness, consistency, and format variations across all sites. This often reveals unexpected silos and data quality issues that require significant upfront remediation.
Define Data Model & Integration Strategy: Design a target data model that consolidates disparate data into a harmonized format (e.g., FHIR-compliant structures for clinical data, common data models for genomics). Plan a phased data ingestion strategy (batch initially, then real-time streaming).
Architect Secure Data Pipelines (HIPAA/GDPR Compliant):
1. Ingestion: Utilize robust ETL/ELT tools (e.g., Apache NiFi, Azure Data Factory, AWS Glue) or custom microservices for pulling data from source systems.
2. Transformation: Implement data cleaning, normalization, de-identification (if necessary), and feature engineering logic (e.g., converting unstructured text to structured disease markers). Use Python with libraries like pandas, NLTK, spaCy for complex text processing.
3. Storage: Establish a secure, scalable data lake (e.g., Azure Data Lake, AWS S3, Google Cloud Storage) for raw data, and a secure data warehouse for processed, analytics-ready data.
4. Access Controls: Implement granular role-based access control (RBAC), encryption at rest and in transit, and robust audit trails.
Vendor Selection & PoC (Proof of Concept): Evaluate AI platforms (e.g., IBM Watson for Oncology, Tempus, Google Health, custom LLM solutions). Conduct a PoC with a small, representative dataset and a limited scope to validate technical feasibility and initial clinical utility for AI treatment planning.

3. AI Model Configuration & Initial Clinical Contextualization (Phase 2)

Knowledge Base Customization: Integrate your institution's specific clinical pathways, local formularies, guidelines (e.g., NCCN, ASCO), and unique patient population characteristics into the chosen AI's knowledge base. This is crucial for relevant AI treatment planning.
Prompt Engineering & Output Tailoring: Work closely with oncologists to refine the prompts given to the AI and customize the format and content of its output (e.g., emphasizing certain metrics, prioritizing specific types of evidence, linking to internal resources). Iterate on this until the output is clinically actionable and user-friendly.
Define Performance Metrics & Evaluation Framework: How will you measure success beyond anecdotal feedback? Define quantitative clinical (e.g., protocol adherence, outcome surrogate metrics), operational (e.g., time savings, cost reduction), and user adoption metrics. Establish a baseline before deployment.
Pilot Program Design: Select a limited pilot group (e.g., 5-10 oncologists in one department or clinic) and clearly define success criteria, training protocols, and a feedback mechanism. Train users thoroughly on how to use the AI tool, interpret its recommendations, and provide structured feedback.

4. Workflow Integration & Scaled Deployment (Phase 3)

Seamless EHR Integration: Design and implement the integration pathway (e.g., SMART on FHIR application, API proxy service) to launch the AI tool directly from the patient chart within your EHR. Ensure single sign-on (SSO) and minimal context switching for clinicians.
Rollout Strategy: Plan a phased rollout to the broader network, starting with sites showing high enthusiasm and good data quality. Provide ongoing training and dedicated technical and clinical support at each stage.
Continuous Monitoring & Feedback Loop:
- Technical Monitoring: Monitor API latency, uptime, data pipeline health, and system performance using tools like Datadog, Prometheus, or Splunk.
- Clinical Monitoring: Track defined metrics via dashboards (e.g., Tableau, Power BI). Collect structured feedback from clinicians on recommendation agreement, reasons for deviation, and perceived utility.
- Model Retraining & Refinement: Establish a regular cadence for model updates and retraining based on new research, updated guidelines, and your institution's anonymized outcome data. This is critical for sustained value in AI treatment planning.

5. Governance & Ethical Oversight (Ongoing)

Data Governance Committee: Maintain an active committee to ensure ongoing data quality, privacy compliance, and appropriate data utilization for AI model improvement.
AI Ethics Review Board: Establish a body to regularly review the AI's performance for bias, fairness, and potential unintended consequences. This might involve clinical ethicists, patient advocates, and technical experts.
Transparency & Explainability: Continuously strive to improve the AI's ability to explain its reasoning. For regulatory compliance and clinician trust, "black box" solutions are generally unacceptable in a clinical setting, especially for critical decisions like AI treatment planning.

AI Treatment Planning: Optimize Oncology Protocols with IBM Watson Health is ideal for teams that need faster execution and measurable outcomes.