Convergent AI in Reducing Overdiagnosis, Overtreatment, and Misdiagnosis
The current healthcare situation:
High-cost/low-quality healthcare is now a national crisis. The government is trying to make it right but is often making it worse.
The problem is partly socioeconomic; addressing:
- chronic-care management
- patients as consumers
- patient-physician engagement
- population health
Meanwhile, micro/macro economics and technology have come together to create a data boom in healthcare. The goal is to create:
- value-based care
- stricter regulations
- decreased payments
- improved outcomes and quality
The biggest roadblock to success: waste
Waste costs the healthcare system between $750 billion – $935 billion annually. This is close to ⅓ of all healthcare costs.
- Administrative complexity: $265.6 billion
- Pricing failure: $230.7 billion — $240.5 billion
- Overtreatment: $75.7 billion — $101.2 billion
- Fraud and abuse: $58.5 billion — $83.9 billion
- Failure of care coordination: $27.2 billion — $78.2 billion
- Failure of care: $102.4 billion — $165.7 billion
Artificial intelligence (AI) may offer concrete solutions:
The more artificial intelligence (AI) techniques are integrated into overdiagnosis, overtreatment and misdiagnosis solutions, the better it works.
Problem: we often don’t know how to integrate the data; the compression of data is huge and cumbersome.
Types of data that may benefit from AI and ML:
- Social media
- Drug profile
The following case studies illustrate how AI and ML may deliver industry solutions:
Case Study: breast cancer risk assessment and overdiagnosis
- 23 million mammograms are done annually in the U.S.
- 266,600 new cases of invasive breast cancer are diagnosed annually in the U.S.
- False positive rate from mammograms is estimated to be 7% to 10 %
- False positive mammograms are estimated to cost $4 billion annually.
- Patients who have received false-positive diagnoses have shown to exhibit a higher level of anxiety and lower level of self-esteem.
A breast cancer diagnosis is based on a risk standard seen in the imaging. Many of these cases will go to biopsy, with false positives at 85 percent. What that means is that a biopsy may have been performed for nothing, and it ultimately costs the country about $4 billion every year.
Possible solution: An intelligent-augmented breast cancer risk calculator (iBRISK) could hopefully give a better score than our current model. How? Using convergent AI, which includes:
- natural language processing
- image analysis
- deep learning
- data mining
- big, multi-modal BI-RADS patient data.
How it works: key in the data and a recommendation is given. Doctors and patients can still consider a biopsy if they want.
Results: so far, 14,000 patients were tested. As a result, about 80 percent of women do not have to go for a biopsy. This could possibly save the country about $3 billion annually.
The most vital benefit of this innovation is actually psychological — avoiding damage to the well-being of the women who would otherwise endure an unnecessary biopsy procedure. This component may be unmeasurable.
AI may help with other screenings — such as those for lung cancer — by making them more precise and reducing waste.
Another example, this time addressing misdiagnosis:
Case study: Misdiagnosis of Acute Ischemic Stroke
- 800,000 stroke incidents in the U.S. annually
- 80 percent of strokes are acute ischemic strokes
- Challenges to acute stroke detection: difficult to identify early ischemic changes, low detection rate of early ischemic change
- Computed tomography (CT ) is widely available in clinical sites and emergency rooms. It’s difficult to identify the changes.
Diffusion MRI, the acceptable gold standard, can detect stroke very early but is often not performed due to time and cost restraints. Usually, this is only used when a patient is already in the hospital.
How AI can help in terms of early detection:
AI allows us to understand the patient’s data. Doctors may not be able to detect a stroke, but they can understand relationships: tissue-specific non-linear relationships from apparent diffusion coefficient (ADC) to CT density. A map can be generated from the CT scan.
Status: Clinical trials are currently being performed.
Case study: facial recognition and voice interpretation via deep learning:
By studying videos of patients reading stories and using language, stroke symptoms and diagnoses can be better understood. Although the technology is not yet perfected, this may be enough for the AI to learn and predict more accurately.
Status: clinical trials are currently being performed on about 1,000 cases.
Bottom line: Accelerate the pace of delivery system transformation for patient, purchaser, provider and payer:
- Save costs
- Improve operation
- Improve healthcare