AI use in cancer treatment is in the early stages
Machine learning holds promise in identifying similar patients and best practices, says Ayan Bjattacharya.
Artificial intelligence holds great promise for improving healthcare delivery through automation and machine learning that leads to better care paths and streamlined process efficiencies. But to date, the industry has promised much but delivered little, even as users have high expectations that demand detailed diagnostics and evidence of consistent results.
A panel of institutional, consulting and industry life science specialists discussed the state of AI in cancer treatment during the MedCity Converge conference in Philadelphia. Their observations, while positive in outlook, dealt more with traditional challenges of automation and data integration, and mostly not ready-for-market niche solutions.
"The current use case for artificial intelligence and machine learning is automate data processing so we can identify similar patients and gather some best practices and guideposts for treatment," says Ayan Bjattacharya of Deloitte Consulting. "Then, we look at how to optimize the whole process with AI to break silos of data."
Oncology providers can greatly benefit from machine-discovered markers and indicators that identify patients suitable for specialized care, which helps to compensate for time and labor constraints they face. "We want to find the right patient who can specifically benefit from our care because there are lots of different providers out there," says Tulia Haddad, chair of Breast Medical Oncology at Mayo Clinic. "With the right profiles, we can quickly identify the patients who can benefit."
Also See: Machine learning predicts 1-year mortality using EHR data
Tools are in development for cognitive systems that can identify key attributes of cancer to help identify cancers and treatments. But tool use faces a bottleneck in the very traditional challenges of data granularity and integration required to make machine learning useful.
John Quackenbush, director of the Center for Cancer Computational Biology at Dana-Farber Cancer institute, recalled one promising project that failed simply because the data was not uniformly defined. "We used Fit Bit data combined with other monitoring devices, where the dataset was very large but was totally incompatible. So in the end, it wasn't a big data problem; it was a messy data problem."
However, Quackenbush says AI and machine learning is very applicable to particular pathologies. So with the appropriate constraints, such as tumor weight, dimensions or physical location, an oncologist can more quickly identify and direct individual patient care.
The more broadly evidence is collected, the better such a system can work, panelists said. Remote monitoring is also a valuable input for AI, says Haddad, where predictors such as diet or time spent in bed can lead to adjustments that reduce recovery time. But personalized care remains an automation problem because there are too few such examples to train the system. This training also extends to humans whose role is to statistically consider whether medication, history or personal behavioral data is valuable or just a meaningless correlation. "Everything here happens in careful advances because the stakes are higher," Quackenbush says.
The current high ground of AI and machine learning is giving caregivers information that otherwise would take days to sift through manually, but this requires data from more than just one hospital or health system to create a dataset that is clinically valuable.
"For every hour you spend with a patient, you spend two hours behind a computer, and that's the promise of AI," says Carla Liebowitz, head of corporate development at cloud-based AI imaging specialist Arterys. "When I say this is coming, there is immense skepticism because technology got us into this problem in the first place. But the hope is to use data mining and data extraction to do the work for us so the physician can make the best decision and develop a care plan."
A panel of institutional, consulting and industry life science specialists discussed the state of AI in cancer treatment during the MedCity Converge conference in Philadelphia. Their observations, while positive in outlook, dealt more with traditional challenges of automation and data integration, and mostly not ready-for-market niche solutions.
"The current use case for artificial intelligence and machine learning is automate data processing so we can identify similar patients and gather some best practices and guideposts for treatment," says Ayan Bjattacharya of Deloitte Consulting. "Then, we look at how to optimize the whole process with AI to break silos of data."
Oncology providers can greatly benefit from machine-discovered markers and indicators that identify patients suitable for specialized care, which helps to compensate for time and labor constraints they face. "We want to find the right patient who can specifically benefit from our care because there are lots of different providers out there," says Tulia Haddad, chair of Breast Medical Oncology at Mayo Clinic. "With the right profiles, we can quickly identify the patients who can benefit."
Also See: Machine learning predicts 1-year mortality using EHR data
Tools are in development for cognitive systems that can identify key attributes of cancer to help identify cancers and treatments. But tool use faces a bottleneck in the very traditional challenges of data granularity and integration required to make machine learning useful.
John Quackenbush, director of the Center for Cancer Computational Biology at Dana-Farber Cancer institute, recalled one promising project that failed simply because the data was not uniformly defined. "We used Fit Bit data combined with other monitoring devices, where the dataset was very large but was totally incompatible. So in the end, it wasn't a big data problem; it was a messy data problem."However, Quackenbush says AI and machine learning is very applicable to particular pathologies. So with the appropriate constraints, such as tumor weight, dimensions or physical location, an oncologist can more quickly identify and direct individual patient care.
The more broadly evidence is collected, the better such a system can work, panelists said. Remote monitoring is also a valuable input for AI, says Haddad, where predictors such as diet or time spent in bed can lead to adjustments that reduce recovery time. But personalized care remains an automation problem because there are too few such examples to train the system. This training also extends to humans whose role is to statistically consider whether medication, history or personal behavioral data is valuable or just a meaningless correlation. "Everything here happens in careful advances because the stakes are higher," Quackenbush says.
The current high ground of AI and machine learning is giving caregivers information that otherwise would take days to sift through manually, but this requires data from more than just one hospital or health system to create a dataset that is clinically valuable.
"For every hour you spend with a patient, you spend two hours behind a computer, and that's the promise of AI," says Carla Liebowitz, head of corporate development at cloud-based AI imaging specialist Arterys. "When I say this is coming, there is immense skepticism because technology got us into this problem in the first place. But the hope is to use data mining and data extraction to do the work for us so the physician can make the best decision and develop a care plan."
More for you
Loading data for hdm_tax_topic #better-outcomes...