番茄社区

Between low-dose CT lung cancer screening and the wide range of treatment options available, you would think the playing field for lung cancer would be more level. But in the time period between 1990 and 2019 more men in the U.S. experienced a significant decline in lung cancer incidence than women did (37.73% vs 1.41%), reported Vamsidhar Velcheti, MD, of the Perlmutter Cancer Center at New York University in New York City, and colleagues.

Similarly, for a decrease in lung cancer mortality, men again came out ahead in those 30 years (40.23% vs 6.01%).

What might be going on here? In their study in , Velcheti's group noted that an analysis of U.S. lung cancer histology through 2010 showed that the rates of squamous, large-cell, and small-cell carcinomas have all been on a downward trajectory across all sexes, but that the rates of adenocarcinoma have remained relatively steady in males while increasing in females -- particularly young females -- among all racial and ethnic groups.

While the incidence and mortality disparities were prominent when the analysis separated the data for males from that for females, there was some general good news: a meaningful decline in lung cancer incidence overall (23.35%) in the U.S., and disability-adjusted life years also decreased by 35.94% for both sexes combined, the researchers reported.

In addition, mortality-to-incidence (MII) indices decreased overall, although there were variations by state, with New Mexico, Oklahoma, and Utah not seeing any positive shifts, for instance.

Velcheti and co-authors stressed that the strength of their analysis was based on "trends rather than absolute annual mortality rates," which "allows for the assessment of population-level trends over an extended observation period using the annual mortality data collected from the GBD [Global Burden of Disease database]."

The investigators also conceded that since the study was limited by its observational nature, it was not possible to conclude causal inferences, or account for certain potential confounders despite using gender-specific and age-standardized incidence and mortality rates.

The value of the findings is that they highlight the positive: "MIIs decreased in all states, probably because of a decrease in incidence and advancements in treatment" as well as the not-so-positive -- specifically the increasing numbers in most states for females, which call for more exploration, the researchers said.

In related research, Marshall H. Chin, MD, MPH, of the University of Chicago, and colleagues explained that healthcare algorithms are used for diagnosis, treatment, prognosis, risk stratification, and allocation of resources. Bias in the development and use of algorithms, the researchers said, "can lead to worse outcomes for racial and ethnic minoritized groups and other historically marginalized populations such as individuals with lower income."

Chin's group was part of an Agency for Healthcare Research and Quality-National Institute on Minority Health and Health Disparities panel that created a "conceptual framework to apply guiding principles across an algorithm's life cycle, centering health and health care equity for patients and communities as the goal, within the wider context of structural racism and discrimination," the researchers explained in .

In an accompanying invited , Adrian F. Hernandez, MD, MHS, of the Duke Clinical Research Institute in Durham, North Carolina, and colleagues explained that these guiding principles "add precision and specificity that go beyond broad mandates, such as the White House Blueprint for an AI [artificial intelligence] Bill of Rights and National Institute of Standards and Technology AI Risk Management Framework."

This type of guidance may well be particularly meaningful in lung cancer care, in -- using machine learning, deep learning, and radiomics to detect and characterize lung nodules -- and in disease , as well as predicting .

In a , Chin, along with co-author Lucila Ohno-Machado, MD, PhD, MBA, of Yale School of Medicine in New Haven, Connecticut, shared more details about the framework.

What was the impetus for the study?

Chin: Healthcare algorithms -- these are mathematical models that inform decision-making in both clinical care as well as healthcare in general -- have great potential to improve outcomes as well as potential to do harm. We know from other industries, such as banking, housing, education, that there is evidence of racial and ethnic bias in these algorithms, and there are many examples of racial and ethnic bias in healthcare algorithms.

Ohno-Machado: And related to AI, are there models that learn from data? So if the data are biased themselves, it's very possible that the algorithms themselves, the AI models, will become biased. I think there is much more attention now that AI has become popular on racial and ethnic disparities, although they were there all the time.

How do the guiding principles address the "?"

Chin: That is the beginning to end cycle -- first coming up with the problem the algorithm is designed to solve, then finding the data to develop the algorithm, actually developing the algorithm, deploying it in the field, and then monitoring the effects. There can be bias that enters in both the data entry and all the different aspects of the algorithm life cycle, so we need to think about potential biases throughout the entire algorithm life cycle.

The article states that "technical definitions and metrics of fairness often do not translate clearly or intuitively to ethical, legal, social, and economic conceptions of fairness." Would you elaborate on that?

Chin: The bottom line is that biases can affect patient outcomes as well as fairness in the way that resources are allocated to different patients and communities. From a technical standpoint, there are a variety of ways to measure, and then try to address, algorithmic bias.

These basically have to do with tradeoffs between trying to maximize the overall accuracy of an algorithm across all patients, and minimizing the differences in the accuracy in the algorithmic program across different populations.

Ohno-Machado: One way to think about this is that you are trying to minimize the overall error for the whole population. However, if you're making mistakes consistently in a smaller portion of the population, that needs to be considered, because you might be unfair to that population while you are minimizing the error overall.

Chin: This is one reason that ensuring algorithmic fairness needs to be a team sport. There needs to be collaboration among different stakeholders, including the technical experts who develop the algorithms, the users of the algorithms, patients, and the communities. Everyone has an important perspective that needs to be brought to the table.

Read the study here.