Calling Dr. DeepSeek, Stat!
How AI Can Assist Physicians in Delivering Better Patient Care
Introduction
DeepSeek has recently exploded into the public consciousness for its groundbreaking AI assistant, which rivals leading models like OpenAI o1. Notably, DeepSeek achieved this with substantially lower costs and resource requirements, challenging the prevailing notion that advanced AI development necessitates massive investments. This efficiency has the potential to democratize AI applications, making them more accessible across various sectors, including healthcare.
As an emergency medicine physician, I often operate under intense pressure and time constraints. Although I, along with other ER physicians, thrive in these tough conditions, we will occasionally miss important diagnoses. Most recently, I had one of those cases where I thought that AI tools like DeepSeek might have been able to help me not miss that diagnosis. With the capabilities of AI models like DeepSeek to advance rapidly and lower costs, this potentially can reduce barriers to access powerful healthcare applications to physicians like myself and ultimately lead to greater access to high-quality care for patients.
So I thought to myself: “Let’s see what DeepSeek would do with this case that I missed!”
A Case of Missed Diagnosis
I was finishing a grueling 9-hour shift, seeing over 35 patients. Toward the end of my shift, I encountered a young female in her 20s, 10-14 days after a vaginal delivery. She presented with abdominal pain and fever. While she had reported a fever at home, she was afebrile in the ER. She had lower abdominal tenderness, possible burning with urination, mild vaginal bleeding, and a heart rate in the low 100s. Her blood pressure was also normal.
Given her symptoms, I considered the typical differential diagnoses for fever and lower abdominal pain. I ordered labs, a urinalysis, and a CT scan to rule out appendicitis.
Lab Results: White Blood Cell count elevated at 14
Urinalysis: Leukocyte esterase 1+, nitrite negative, WBC 10-25, minimal bacteria
CT Scan: No acute findings, including a normal appendix
Cognitive Biases and the Missed Diagnosis
Seeing a negative CT scan and a urinalysis suggestive of a urinary tract infection (UTI), I concluded that a complicated UTI explained the elevated WBC. The patient’s condition improved while in the ER, and her heart rate normalized. I discharged her with antibiotics for a UTI.
However, while finishing my notes at the end of my shift, I realized I might have missed the actual diagnosis: endometritis. In retrospect, I recognized that I had fallen into two common cognitive biases:
Availability Heuristic: I anchored on the most available information—her urinalysis—without considering the broader clinical picture.
Recency Bias: Given the fatigue from a long shift, I was more likely to diagnose a common condition I had seen frequently that day and undoubtedly, I saw my fair share of patients with UTI earlier in the shift
Realizing my mistake, I called the patient back for proper treatment of endometritis.
How AI Could Help
Curious about how AI might assist in such cases, I inputted this case into DeepSeek-R1. The results were eye-opening.
DeepSeek’s reasoning immediately flagged endometritis as a key diagnosis, stating: “Endometritis is a big one, right?”
This was impressive because it aligned with the correct diagnosis that I had initially overlooked due to fatigue and bias.
DeepSeek also considered other less likely diagnoses and even very unlikely ones like mastitis. Ultimately, it correctly reasoned away from those diagnoses.
Additionally, it acknowledged my thought process for ordering the CT scan: “Appendicitis? That can happen postpartum.”
When I introduced the urinalysis and CT scan results, I wondered if DeepSeek would fall into the same cognitive bias I had. Instead, it correctly noted that “endometritis is primarily a clinical diagnosis” and questioned whether the urinalysis findings were truly indicative of a UTI.
It even identified the urinalysis as a possible red herring—a level of clinical reasoning that mirrored how I would think through a case with more time and less fatigue.
For comparison, I tested the same prompt with OpenAI’s o1, and it reached similar conclusions, suggesting that AI has the potential to serve as a valuable second opinion in clinical decision-making.
The Future of AI in Medicine
Keep in mind I fed the digested information to DeepSeek and it was not getting the largely messy, unstructured information from the patient that I received. What is impressive is the potential to help me find diagnoses that I could miss. This is not about replacing doctors. AI is not taking my job. Rather, AI has the potential to serve as a powerful assistant, helping physicians mitigate human limitations such as fatigue and cognitive biases.
We are already witnessing AI’s impact in clinical documentation, with companies like Abridge and Suki reducing administrative burdens for doctors. Studies show that AI-powered documentation tools save clinicians several hours per week1, allowing them to spend more time with patients.
The next step? AI assistants that help physicians catch missed diagnoses. Research suggests that diagnostic errors occur in 10-15% of cases2, with fatigue and cognitive biases playing a significant role. AI could serve as an invaluable second set of eyes, ensuring that critical diagnoses are not overlooked.
As Dr. Anthony Chang, Chair of the American Board of AI in Medicine (ABAIM) and Chief Intelligence and Innovation Officer at Children's Hospital of Orange County (CHOC), famously said in one of his wonderful AI courses I attended: “AI is not going to replace the radiologist. A radiologist using AI will replace the radiologist not using AI.”
The same applies to all of medicine. AI won’t replace doctors—but doctors using AI will provide safer, more accurate, and more efficient care than ever before. At Good AI Capital, we are excited to see the future of AI in improving healthcare and working alongside startups who are using AI to tackle the multitude of healthcare challenges.
References
"How AI Assistants Like Abridge and Suki Save Physicians Time." Business Insider, September 2024, https://www.businessinsider.com/doctor-ai-assistant-saves-time-reduces-stress-2024-9.
"Diagnostic Errors in the Emergency Department: A Systematic Review," Agency for Healthcare Research and Quality, 2022. Link ↩
So insightful! Thank you for sharing!