Fitbit Data & AI Predict Surgery Outcomes, Mental Health – & COVID-19’s Impact

The effort to predict how patients will fare after major surgery, increasingly reliant on machine learning and data from wearable devices, hit an unexpected snag: the COVID-19 pandemic. Researchers at Washington University in St. Louis (WashU) discovered that their models, designed to forecast outcomes following pancreatic surgery, were thrown off by factors related to the pandemic, highlighting the complex interplay between public health crises and even highly individualized medical predictions.

The Unexpected Variable

WashU researchers have been integrating data from consumer-grade Fitbit wristbands into machine-learning models. The goal is to improve predictions around surgical outcomes, post-operative pain, and potential mental health challenges. These models analyze activity levels, sleep patterns, and other physiological data to provide a more holistic picture of a patient’s recovery trajectory. However, when applied to pancreatic surgery patients, the models began to produce inaccurate forecasts. The culprit? The COVID-19 pandemic. The initial operate focused on pancreatic surgery, but the approach is being expanded to other procedures and conditions. WashU’s Department of Surgery offers specialized care for a range of conditions, including cancers of the pancreas, liver, gallbladder, bile duct, stomach, and duodenum, as well as non-malignant issues like gallstones and gastroesophageal reflux. More information about the department and its providers is available on their website.

Pandemic Impacts on Surgical Care

The pandemic’s influence wasn’t simply a matter of patients being sicker with COVID-19. A study published in PubMed in February 2024, analyzing data from the National Cancer Database between 2017 and 2020, revealed a more nuanced picture. The research showed a decrease in the number of pancreatic ductal adenocarcinoma (PDAC) diagnoses in 2020 compared to the preceding years. This suggests that fewer people were diagnosed with the cancer during the height of the pandemic, potentially due to disruptions in healthcare access and screening programs. The study found a decrease in the percentage of patients receiving surgery or radiotherapy in 2020.

Perhaps more surprisingly, even the suspicion of COVID-19 – even in patients who tested negative – lengthened the time from diagnosis to surgery. The median time from diagnosis to surgery increased significantly from 34 days in 2017 to 81 days in 2020. Patients who tested positive for COVID-19, even if only briefly, experienced even longer delays, with a median time of 140 days from diagnosis to surgery, compared to 112 days for those who tested negative. This suggests that hospitals were prioritizing resources and postponing elective surgeries to manage the influx of COVID-19 patients, and that concerns about infection risk were influencing clinical decision-making.

What Does This Mean for Predictive Models?

The WashU researchers’ experience underscores a critical limitation of many machine-learning models in healthcare: they are often trained on historical data that may not accurately reflect current or future conditions. The pandemic represented a significant disruption to healthcare systems and patient behavior, creating a situation that the models hadn’t “seen” before. This resulted in inaccurate predictions, highlighting the need for continuous model retraining and adaptation. The study in PubMed confirms that while the quality of the surgery itself didn’t decline during the pandemic, delays in diagnosis and treatment had a measurable impact on patient care.

Beyond Pancreatic Cancer: Broader Implications

The lessons learned from this experience extend beyond pancreatic cancer surgery. The pandemic impacted healthcare across the board, leading to delays in screenings, treatments, and routine care for a wide range of conditions. Any predictive model that relies on historical data from this period may be similarly affected. What we have is particularly relevant for conditions where early diagnosis and treatment are crucial, such as cancer and cardiovascular disease. The impact of the pandemic on surgical AI forecasts, as reported by Medical Xpress, serves as a cautionary tale for the field of artificial intelligence in medicine.

The Role of Wearable Technology and Data

Despite the challenges posed by the pandemic, the use of wearable technology and data-driven models holds significant promise for improving surgical outcomes. Fitbit data, for example, can provide valuable insights into a patient’s baseline activity level, sleep quality, and physiological responses to stress. This information can be used to identify patients who are at higher risk of complications or who may benefit from more intensive post-operative care. However, it’s crucial to acknowledge the limitations of this data. Wearable devices are not medical-grade instruments, and the data they collect may be subject to inaccuracies or biases. Access to wearable technology is not equitable, and relying solely on this data could exacerbate existing health disparities.

What Comes Next: Model Refinement and Surveillance

The WashU researchers are now working to refine their models to account for the impact of the pandemic. This involves incorporating new data from the pandemic period and developing algorithms that are more robust to unexpected disruptions. Ongoing surveillance of surgical outcomes and treatment patterns is also essential to identify any lingering effects of the pandemic and to ensure that patients are receiving the best possible care. Hospitals and healthcare systems are also reviewing their protocols for managing elective surgeries during public health emergencies to minimize delays and ensure equitable access to care. Further research is needed to understand the long-term consequences of pandemic-related care delays and to develop strategies for mitigating these effects. The National Cancer Database will continue to be a valuable resource for tracking trends in cancer diagnosis and treatment, and for identifying areas where improvements are needed.