Speaker

Biography

Abstract

Kirsty Hicks
(GSK)

Kirsty is a Senior Statistics Director at GSK, having joined in 2001 from Roche where she worked as a late phase statistician for a number of years. Throughout her time at GSK Kirsty has supported all gastrointestinal assets in the early phase portfolio covering diseases such as ulcerative colitis and irritable bowel syndrome; supported numerous immunological-inflammatory assets across a wide range of diseases; and more recently heads up the Cancer Epigenetics Biostatistics group. Throughout this time Kirsty has managed an ever growing team of statisticians in addition to being heavily involved in numerous non project initiatives. More recently these have included exploring experimental medicine analysis methods; facilitating prior elicitations and leading the biostatistics team investigating the incorporation and analysis of complex digital data into clinical trials. Kirsty also currently leads the UK Oncology Biostatistics team, working closely with Oncology Biostatistics teams in the US, India, Japan and China.

Advanced Analytics of Digital Data: A focus on sensor data.

What is the evolving area of Digital data you may ask? Nowadays most of us will be wearing a watch monitoring some aspect of our movement or use phone apps that collect health information. The data collected can be referred to as digital biomarkers, and this data, in particular sensor technology is becoming a common feature in clinical trials.

Digital data can be used to collect information on sleep patterns, respiration rate, step count and continuous monitoring of heart rate and energy expenditure. Collecting data through digital devices also increases patient engagement and can provide real time compliance monitoring. The regulations for use of digital technologies in clinical trials are still evolving, and current recommendations for the analysis of such data are limited.

Utilising sensors, and collecting actigraphy data, is a non-invasive method of monitoring activity. An actigraph sensor is worn for a time period (eg a week or more) to measure activity. The movements the actigraph undergoes are continually recorded; where 300 data points per second can be collected. As a statistician initial questions that spring to mind include ‘How can this volume of data be analysed or represented visually?’ Obvious summary measures would include average step count over a day; or the percentage of time a patient is active or sedentary. But a lot of information is lost, what about the raw data? For example, the energy used at the minute level or even second level? Statistical methods that can be applied to this data are under investigation and this presentation will introduce some of these.

Edoardo Lisi
(GSK)

Edoardo Lisi is a Clinical Statistician based in the UK. He has been at GSK since 2018, working in Phase 1 and Phase 2 studies. He is a member of GSK’s Advanced Analytics for Digital Data (AADD) group, where he is focused on the statistical analysis of physical activity data from wearable devices.

Using Generalised Additive Models (GAMs) to analyse wearable data.

Many diseases impact the ability of patients to perform physical activity. The use of digital wearable devices equipped with sensors that continuously collect actigraphy data in clinical studies represents a great opportunity to better understand patient populations and how diseases limit their mobility. However, wearable devices produce vast amounts of data (e.g. second-by-second or minute-by-minute over several weeks). The statistical analysis of such type of data to assess the impact of an intervention is challenging. Ideas proposed in the literature suggest to summarise physical activity through indices of total activity or indices of activity fragmentation. However, there is no consensus yet in the scientific community with respect to how to best analyse actigraphy data. Producing numerical indices condensing long minute-by-minute time-series into one single value may lead to loss of information. Here we describe using generalised additive models (GAMs) to analyse the entire time series. We apply GAMs to the analysis of minute-by-minute activity data in a real case study and show how GAMs can account for the different sources of variation specific to this type of data (e.g. time of the day, weekday vs. weekend) thus better exploiting the information contained in the actigraphy data. We also illustrate how GAMs can help to detect differences not only in volume of activity but also in the shape, i.e. in how the activity spreads over the day.

Munshi Imran Hossain
(Cytel)

Munshi Imran Hossain is an Associate Consultant in the Quantitative Strategies & Data Science group at Cytel. Imran has about 9 years of experience, working in the areas of software development and data science. He is a biomedical engineer by training and has interests in the processing and analysis of biomedical data including data from sensors, gene expression data and protein biomarker data among others. He is also a member of PSI's Special Interest Group for Data Science.

Processing and Analysis of Accelerometer Data in Subjects with Neurodegenerative Disorders.

Today, sensors have become an essential component of the healthcare ecosystem. They generate valuable data that can be processed to create insights to help improve healthcare. Wearable sensors take this a step further by allowing these systems to be "worn" by subjects. This allows for continuous monitoring and hopefully, timely alerts.

The challenge, however, is to process the vast stream of data generated from sensors and extract useful information. This is an involved exercise needing signal processing and statistical expertise. In this case study, I will present the processing and analysis of data generated from an accelerometer. The objective is to be able to detect difficulty in a physiological process among subjects with neurodegenerative disorders. I will give a glimpse into the kinds of features that can be extracted from raw accelerometer data which, in turn, can be used to build machine learning models.

Andrew Potter (CDER, FDA)

Dr. Andrew Potter is a mathematical statistician in the Division of Biometrics I in CDER supporting the review work in the Division of Psychiatry.  His research interests include the use of digital health technologies in clinical trials and the analysis of high frequency outcome data and in involved in working groups at FDA on this topic.  He received his PhD in Biostatistics from the University of Pittsburgh.

With recent developments in wearable biosensors and portable electronics (collectively referred to as digital health technology – DHT) , many clinical trials propose measuring endpoints using a DHT.  DHTs record data at a higher frequency than traditional clinician/patient observed data.  For example, a wearable accelerometer records data multiple times a second every day during follow-up (lasting multiple weeks).  Clinical endpoints are derived from this high frequency data.  Examples include:

• Trials where an experimental drug is hypothesized to improve exercise capacity with the endpoint of daily time spent at or above a specified exercise intensity
• Trials where daily sleep parameters are measured by a one or more sensors

To estimate and test treatment effects measured in these cases, several statistics issues must be addressed.  Unlike current endpoint measurements, DHTs provide endpoint measurements for every day worn.  In the case that different days (e.g., weekends vs. weekdays) are combined to form endpoints, treatment effect estimates may be biased or miss important disease features if days are not exchangeable.  With the greater amount of data collected, missing data can occur in multiple ways (e.g., a missing observation within a day, missing a day within a week, monotone dropout).

This talk discusses potential issues arising in exercise endpoints if a DHT is worn for different times each day during a study.  It also discusses a case study exploring how different statistical analyses of a clinical trial for a new insomnia drug are affected by missing data to motivate discussion of statistical considerations in studies using these new data collection tools.

Speaker

Biography

Abstract

Kirsty Hicks
(GSK)

Kirsty is a Senior Statistics Director at GSK, having joined in 2001 from Roche where she worked as a late phase statistician for a number of years. Throughout her time at GSK Kirsty has supported all gastrointestinal assets in the early phase portfolio covering diseases such as ulcerative colitis and irritable bowel syndrome; supported numerous immunological-inflammatory assets across a wide range of diseases; and more recently heads up the Cancer Epigenetics Biostatistics group. Throughout this time Kirsty has managed an ever growing team of statisticians in addition to being heavily involved in numerous non project initiatives. More recently these have included exploring experimental medicine analysis methods; facilitating prior elicitations and leading the biostatistics team investigating the incorporation and analysis of complex digital data into clinical trials. Kirsty also currently leads the UK Oncology Biostatistics team, working closely with Oncology Biostatistics teams in the US, India, Japan and China.

Advanced Analytics of Digital Data: A focus on sensor data.

What is the evolving area of Digital data you may ask? Nowadays most of us will be wearing a watch monitoring some aspect of our movement or use phone apps that collect health information. The data collected can be referred to as digital biomarkers, and this data, in particular sensor technology is becoming a common feature in clinical trials.

Digital data can be used to collect information on sleep patterns, respiration rate, step count and continuous monitoring of heart rate and energy expenditure. Collecting data through digital devices also increases patient engagement and can provide real time compliance monitoring. The regulations for use of digital technologies in clinical trials are still evolving, and current recommendations for the analysis of such data are limited.

Utilising sensors, and collecting actigraphy data, is a non-invasive method of monitoring activity. An actigraph sensor is worn for a time period (eg a week or more) to measure activity. The movements the actigraph undergoes are continually recorded; where 300 data points per second can be collected. As a statistician initial questions that spring to mind include ‘How can this volume of data be analysed or represented visually?’ Obvious summary measures would include average step count over a day; or the percentage of time a patient is active or sedentary. But a lot of information is lost, what about the raw data? For example, the energy used at the minute level or even second level? Statistical methods that can be applied to this data are under investigation and this presentation will introduce some of these.

Edoardo Lisi
(GSK)

Edoardo Lisi is a Clinical Statistician based in the UK. He has been at GSK since 2018, working in Phase 1 and Phase 2 studies. He is a member of GSK’s Advanced Analytics for Digital Data (AADD) group, where he is focused on the statistical analysis of physical activity data from wearable devices.

Using Generalised Additive Models (GAMs) to analyse wearable data.

Many diseases impact the ability of patients to perform physical activity. The use of digital wearable devices equipped with sensors that continuously collect actigraphy data in clinical studies represents a great opportunity to better understand patient populations and how diseases limit their mobility. However, wearable devices produce vast amounts of data (e.g. second-by-second or minute-by-minute over several weeks). The statistical analysis of such type of data to assess the impact of an intervention is challenging. Ideas proposed in the literature suggest to summarise physical activity through indices of total activity or indices of activity fragmentation. However, there is no consensus yet in the scientific community with respect to how to best analyse actigraphy data. Producing numerical indices condensing long minute-by-minute time-series into one single value may lead to loss of information. Here we describe using generalised additive models (GAMs) to analyse the entire time series. We apply GAMs to the analysis of minute-by-minute activity data in a real case study and show how GAMs can account for the different sources of variation specific to this type of data (e.g. time of the day, weekday vs. weekend) thus better exploiting the information contained in the actigraphy data. We also illustrate how GAMs can help to detect differences not only in volume of activity but also in the shape, i.e. in how the activity spreads over the day.

Munshi Imran Hossain
(Cytel)

Munshi Imran Hossain is an Associate Consultant in the Quantitative Strategies & Data Science group at Cytel. Imran has about 9 years of experience, working in the areas of software development and data science. He is a biomedical engineer by training and has interests in the processing and analysis of biomedical data including data from sensors, gene expression data and protein biomarker data among others. He is also a member of PSI's Special Interest Group for Data Science.

Processing and Analysis of Accelerometer Data in Subjects with Neurodegenerative Disorders.

Today, sensors have become an essential component of the healthcare ecosystem. They generate valuable data that can be processed to create insights to help improve healthcare. Wearable sensors take this a step further by allowing these systems to be "worn" by subjects. This allows for continuous monitoring and hopefully, timely alerts.

The challenge, however, is to process the vast stream of data generated from sensors and extract useful information. This is an involved exercise needing signal processing and statistical expertise. In this case study, I will present the processing and analysis of data generated from an accelerometer. The objective is to be able to detect difficulty in a physiological process among subjects with neurodegenerative disorders. I will give a glimpse into the kinds of features that can be extracted from raw accelerometer data which, in turn, can be used to build machine learning models.

Andrew Potter (CDER, FDA)

Dr. Andrew Potter is a mathematical statistician in the Division of Biometrics I in CDER supporting the review work in the Division of Psychiatry.  His research interests include the use of digital health technologies in clinical trials and the analysis of high frequency outcome data and in involved in working groups at FDA on this topic.  He received his PhD in Biostatistics from the University of Pittsburgh.

With recent developments in wearable biosensors and portable electronics (collectively referred to as digital health technology – DHT) , many clinical trials propose measuring endpoints using a DHT.  DHTs record data at a higher frequency than traditional clinician/patient observed data.  For example, a wearable accelerometer records data multiple times a second every day during follow-up (lasting multiple weeks).  Clinical endpoints are derived from this high frequency data.  Examples include:

• Trials where an experimental drug is hypothesized to improve exercise capacity with the endpoint of daily time spent at or above a specified exercise intensity
• Trials where daily sleep parameters are measured by a one or more sensors

To estimate and test treatment effects measured in these cases, several statistics issues must be addressed.  Unlike current endpoint measurements, DHTs provide endpoint measurements for every day worn.  In the case that different days (e.g., weekends vs. weekdays) are combined to form endpoints, treatment effect estimates may be biased or miss important disease features if days are not exchangeable.  With the greater amount of data collected, missing data can occur in multiple ways (e.g., a missing observation within a day, missing a day within a week, monotone dropout).

This talk discusses potential issues arising in exercise endpoints if a DHT is worn for different times each day during a study.  It also discusses a case study exploring how different statistical analyses of a clinical trial for a new insomnia drug are affected by missing data to motivate discussion of statistical considerations in studies using these new data collection tools.

Speaker

Biography

Abstract

Kirsty Hicks
(GSK)

Kirsty is a Senior Statistics Director at GSK, having joined in 2001 from Roche where she worked as a late phase statistician for a number of years. Throughout her time at GSK Kirsty has supported all gastrointestinal assets in the early phase portfolio covering diseases such as ulcerative colitis and irritable bowel syndrome; supported numerous immunological-inflammatory assets across a wide range of diseases; and more recently heads up the Cancer Epigenetics Biostatistics group. Throughout this time Kirsty has managed an ever growing team of statisticians in addition to being heavily involved in numerous non project initiatives. More recently these have included exploring experimental medicine analysis methods; facilitating prior elicitations and leading the biostatistics team investigating the incorporation and analysis of complex digital data into clinical trials. Kirsty also currently leads the UK Oncology Biostatistics team, working closely with Oncology Biostatistics teams in the US, India, Japan and China.

Advanced Analytics of Digital Data: A focus on sensor data.

What is the evolving area of Digital data you may ask? Nowadays most of us will be wearing a watch monitoring some aspect of our movement or use phone apps that collect health information. The data collected can be referred to as digital biomarkers, and this data, in particular sensor technology is becoming a common feature in clinical trials.

Digital data can be used to collect information on sleep patterns, respiration rate, step count and continuous monitoring of heart rate and energy expenditure. Collecting data through digital devices also increases patient engagement and can provide real time compliance monitoring. The regulations for use of digital technologies in clinical trials are still evolving, and current recommendations for the analysis of such data are limited.

Utilising sensors, and collecting actigraphy data, is a non-invasive method of monitoring activity. An actigraph sensor is worn for a time period (eg a week or more) to measure activity. The movements the actigraph undergoes are continually recorded; where 300 data points per second can be collected. As a statistician initial questions that spring to mind include ‘How can this volume of data be analysed or represented visually?’ Obvious summary measures would include average step count over a day; or the percentage of time a patient is active or sedentary. But a lot of information is lost, what about the raw data? For example, the energy used at the minute level or even second level? Statistical methods that can be applied to this data are under investigation and this presentation will introduce some of these.

Edoardo Lisi
(GSK)

Edoardo Lisi is a Clinical Statistician based in the UK. He has been at GSK since 2018, working in Phase 1 and Phase 2 studies. He is a member of GSK’s Advanced Analytics for Digital Data (AADD) group, where he is focused on the statistical analysis of physical activity data from wearable devices.

Using Generalised Additive Models (GAMs) to analyse wearable data.

Many diseases impact the ability of patients to perform physical activity. The use of digital wearable devices equipped with sensors that continuously collect actigraphy data in clinical studies represents a great opportunity to better understand patient populations and how diseases limit their mobility. However, wearable devices produce vast amounts of data (e.g. second-by-second or minute-by-minute over several weeks). The statistical analysis of such type of data to assess the impact of an intervention is challenging. Ideas proposed in the literature suggest to summarise physical activity through indices of total activity or indices of activity fragmentation. However, there is no consensus yet in the scientific community with respect to how to best analyse actigraphy data. Producing numerical indices condensing long minute-by-minute time-series into one single value may lead to loss of information. Here we describe using generalised additive models (GAMs) to analyse the entire time series. We apply GAMs to the analysis of minute-by-minute activity data in a real case study and show how GAMs can account for the different sources of variation specific to this type of data (e.g. time of the day, weekday vs. weekend) thus better exploiting the information contained in the actigraphy data. We also illustrate how GAMs can help to detect differences not only in volume of activity but also in the shape, i.e. in how the activity spreads over the day.

Munshi Imran Hossain
(Cytel)

Munshi Imran Hossain is an Associate Consultant in the Quantitative Strategies & Data Science group at Cytel. Imran has about 9 years of experience, working in the areas of software development and data science. He is a biomedical engineer by training and has interests in the processing and analysis of biomedical data including data from sensors, gene expression data and protein biomarker data among others. He is also a member of PSI's Special Interest Group for Data Science.

Processing and Analysis of Accelerometer Data in Subjects with Neurodegenerative Disorders.

Today, sensors have become an essential component of the healthcare ecosystem. They generate valuable data that can be processed to create insights to help improve healthcare. Wearable sensors take this a step further by allowing these systems to be "worn" by subjects. This allows for continuous monitoring and hopefully, timely alerts.

The challenge, however, is to process the vast stream of data generated from sensors and extract useful information. This is an involved exercise needing signal processing and statistical expertise. In this case study, I will present the processing and analysis of data generated from an accelerometer. The objective is to be able to detect difficulty in a physiological process among subjects with neurodegenerative disorders. I will give a glimpse into the kinds of features that can be extracted from raw accelerometer data which, in turn, can be used to build machine learning models.

Andrew Potter (CDER, FDA)

Dr. Andrew Potter is a mathematical statistician in the Division of Biometrics I in CDER supporting the review work in the Division of Psychiatry.  His research interests include the use of digital health technologies in clinical trials and the analysis of high frequency outcome data and in involved in working groups at FDA on this topic.  He received his PhD in Biostatistics from the University of Pittsburgh.

With recent developments in wearable biosensors and portable electronics (collectively referred to as digital health technology – DHT) , many clinical trials propose measuring endpoints using a DHT.  DHTs record data at a higher frequency than traditional clinician/patient observed data.  For example, a wearable accelerometer records data multiple times a second every day during follow-up (lasting multiple weeks).  Clinical endpoints are derived from this high frequency data.  Examples include:

• Trials where an experimental drug is hypothesized to improve exercise capacity with the endpoint of daily time spent at or above a specified exercise intensity
• Trials where daily sleep parameters are measured by a one or more sensors

To estimate and test treatment effects measured in these cases, several statistics issues must be addressed.  Unlike current endpoint measurements, DHTs provide endpoint measurements for every day worn.  In the case that different days (e.g., weekends vs. weekdays) are combined to form endpoints, treatment effect estimates may be biased or miss important disease features if days are not exchangeable.  With the greater amount of data collected, missing data can occur in multiple ways (e.g., a missing observation within a day, missing a day within a week, monotone dropout).

This talk discusses potential issues arising in exercise endpoints if a DHT is worn for different times each day during a study.  It also discusses a case study exploring how different statistical analyses of a clinical trial for a new insomnia drug are affected by missing data to motivate discussion of statistical considerations in studies using these new data collection tools.

Speaker

Biography

Abstract

Kirsty Hicks
(GSK)

Kirsty is a Senior Statistics Director at GSK, having joined in 2001 from Roche where she worked as a late phase statistician for a number of years. Throughout her time at GSK Kirsty has supported all gastrointestinal assets in the early phase portfolio covering diseases such as ulcerative colitis and irritable bowel syndrome; supported numerous immunological-inflammatory assets across a wide range of diseases; and more recently heads up the Cancer Epigenetics Biostatistics group. Throughout this time Kirsty has managed an ever growing team of statisticians in addition to being heavily involved in numerous non project initiatives. More recently these have included exploring experimental medicine analysis methods; facilitating prior elicitations and leading the biostatistics team investigating the incorporation and analysis of complex digital data into clinical trials. Kirsty also currently leads the UK Oncology Biostatistics team, working closely with Oncology Biostatistics teams in the US, India, Japan and China.

Advanced Analytics of Digital Data: A focus on sensor data.

What is the evolving area of Digital data you may ask? Nowadays most of us will be wearing a watch monitoring some aspect of our movement or use phone apps that collect health information. The data collected can be referred to as digital biomarkers, and this data, in particular sensor technology is becoming a common feature in clinical trials.

Digital data can be used to collect information on sleep patterns, respiration rate, step count and continuous monitoring of heart rate and energy expenditure. Collecting data through digital devices also increases patient engagement and can provide real time compliance monitoring. The regulations for use of digital technologies in clinical trials are still evolving, and current recommendations for the analysis of such data are limited.

Utilising sensors, and collecting actigraphy data, is a non-invasive method of monitoring activity. An actigraph sensor is worn for a time period (eg a week or more) to measure activity. The movements the actigraph undergoes are continually recorded; where 300 data points per second can be collected. As a statistician initial questions that spring to mind include ‘How can this volume of data be analysed or represented visually?’ Obvious summary measures would include average step count over a day; or the percentage of time a patient is active or sedentary. But a lot of information is lost, what about the raw data? For example, the energy used at the minute level or even second level? Statistical methods that can be applied to this data are under investigation and this presentation will introduce some of these.

Edoardo Lisi
(GSK)

Edoardo Lisi is a Clinical Statistician based in the UK. He has been at GSK since 2018, working in Phase 1 and Phase 2 studies. He is a member of GSK’s Advanced Analytics for Digital Data (AADD) group, where he is focused on the statistical analysis of physical activity data from wearable devices.

Using Generalised Additive Models (GAMs) to analyse wearable data.

Many diseases impact the ability of patients to perform physical activity. The use of digital wearable devices equipped with sensors that continuously collect actigraphy data in clinical studies represents a great opportunity to better understand patient populations and how diseases limit their mobility. However, wearable devices produce vast amounts of data (e.g. second-by-second or minute-by-minute over several weeks). The statistical analysis of such type of data to assess the impact of an intervention is challenging. Ideas proposed in the literature suggest to summarise physical activity through indices of total activity or indices of activity fragmentation. However, there is no consensus yet in the scientific community with respect to how to best analyse actigraphy data. Producing numerical indices condensing long minute-by-minute time-series into one single value may lead to loss of information. Here we describe using generalised additive models (GAMs) to analyse the entire time series. We apply GAMs to the analysis of minute-by-minute activity data in a real case study and show how GAMs can account for the different sources of variation specific to this type of data (e.g. time of the day, weekday vs. weekend) thus better exploiting the information contained in the actigraphy data. We also illustrate how GAMs can help to detect differences not only in volume of activity but also in the shape, i.e. in how the activity spreads over the day.

Munshi Imran Hossain
(Cytel)

Munshi Imran Hossain is an Associate Consultant in the Quantitative Strategies & Data Science group at Cytel. Imran has about 9 years of experience, working in the areas of software development and data science. He is a biomedical engineer by training and has interests in the processing and analysis of biomedical data including data from sensors, gene expression data and protein biomarker data among others. He is also a member of PSI's Special Interest Group for Data Science.

Processing and Analysis of Accelerometer Data in Subjects with Neurodegenerative Disorders.

Today, sensors have become an essential component of the healthcare ecosystem. They generate valuable data that can be processed to create insights to help improve healthcare. Wearable sensors take this a step further by allowing these systems to be "worn" by subjects. This allows for continuous monitoring and hopefully, timely alerts.

The challenge, however, is to process the vast stream of data generated from sensors and extract useful information. This is an involved exercise needing signal processing and statistical expertise. In this case study, I will present the processing and analysis of data generated from an accelerometer. The objective is to be able to detect difficulty in a physiological process among subjects with neurodegenerative disorders. I will give a glimpse into the kinds of features that can be extracted from raw accelerometer data which, in turn, can be used to build machine learning models.

Andrew Potter (CDER, FDA)

Dr. Andrew Potter is a mathematical statistician in the Division of Biometrics I in CDER supporting the review work in the Division of Psychiatry.  His research interests include the use of digital health technologies in clinical trials and the analysis of high frequency outcome data and in involved in working groups at FDA on this topic.  He received his PhD in Biostatistics from the University of Pittsburgh.

With recent developments in wearable biosensors and portable electronics (collectively referred to as digital health technology – DHT) , many clinical trials propose measuring endpoints using a DHT.  DHTs record data at a higher frequency than traditional clinician/patient observed data.  For example, a wearable accelerometer records data multiple times a second every day during follow-up (lasting multiple weeks).  Clinical endpoints are derived from this high frequency data.  Examples include:

• Trials where an experimental drug is hypothesized to improve exercise capacity with the endpoint of daily time spent at or above a specified exercise intensity
• Trials where daily sleep parameters are measured by a one or more sensors

To estimate and test treatment effects measured in these cases, several statistics issues must be addressed.  Unlike current endpoint measurements, DHTs provide endpoint measurements for every day worn.  In the case that different days (e.g., weekends vs. weekdays) are combined to form endpoints, treatment effect estimates may be biased or miss important disease features if days are not exchangeable.  With the greater amount of data collected, missing data can occur in multiple ways (e.g., a missing observation within a day, missing a day within a week, monotone dropout).

This talk discusses potential issues arising in exercise endpoints if a DHT is worn for different times each day during a study.  It also discusses a case study exploring how different statistical analyses of a clinical trial for a new insomnia drug are affected by missing data to motivate discussion of statistical considerations in studies using these new data collection tools.

Speaker

Biography

Abstract

Kirsty Hicks
(GSK)

Kirsty is a Senior Statistics Director at GSK, having joined in 2001 from Roche where she worked as a late phase statistician for a number of years. Throughout her time at GSK Kirsty has supported all gastrointestinal assets in the early phase portfolio covering diseases such as ulcerative colitis and irritable bowel syndrome; supported numerous immunological-inflammatory assets across a wide range of diseases; and more recently heads up the Cancer Epigenetics Biostatistics group. Throughout this time Kirsty has managed an ever growing team of statisticians in addition to being heavily involved in numerous non project initiatives. More recently these have included exploring experimental medicine analysis methods; facilitating prior elicitations and leading the biostatistics team investigating the incorporation and analysis of complex digital data into clinical trials. Kirsty also currently leads the UK Oncology Biostatistics team, working closely with Oncology Biostatistics teams in the US, India, Japan and China.

Advanced Analytics of Digital Data: A focus on sensor data.

What is the evolving area of Digital data you may ask? Nowadays most of us will be wearing a watch monitoring some aspect of our movement or use phone apps that collect health information. The data collected can be referred to as digital biomarkers, and this data, in particular sensor technology is becoming a common feature in clinical trials.

Digital data can be used to collect information on sleep patterns, respiration rate, step count and continuous monitoring of heart rate and energy expenditure. Collecting data through digital devices also increases patient engagement and can provide real time compliance monitoring. The regulations for use of digital technologies in clinical trials are still evolving, and current recommendations for the analysis of such data are limited.

Utilising sensors, and collecting actigraphy data, is a non-invasive method of monitoring activity. An actigraph sensor is worn for a time period (eg a week or more) to measure activity. The movements the actigraph undergoes are continually recorded; where 300 data points per second can be collected. As a statistician initial questions that spring to mind include ‘How can this volume of data be analysed or represented visually?’ Obvious summary measures would include average step count over a day; or the percentage of time a patient is active or sedentary. But a lot of information is lost, what about the raw data? For example, the energy used at the minute level or even second level? Statistical methods that can be applied to this data are under investigation and this presentation will introduce some of these.

Edoardo Lisi
(GSK)

Edoardo Lisi is a Clinical Statistician based in the UK. He has been at GSK since 2018, working in Phase 1 and Phase 2 studies. He is a member of GSK’s Advanced Analytics for Digital Data (AADD) group, where he is focused on the statistical analysis of physical activity data from wearable devices.

Using Generalised Additive Models (GAMs) to analyse wearable data.

Many diseases impact the ability of patients to perform physical activity. The use of digital wearable devices equipped with sensors that continuously collect actigraphy data in clinical studies represents a great opportunity to better understand patient populations and how diseases limit their mobility. However, wearable devices produce vast amounts of data (e.g. second-by-second or minute-by-minute over several weeks). The statistical analysis of such type of data to assess the impact of an intervention is challenging. Ideas proposed in the literature suggest to summarise physical activity through indices of total activity or indices of activity fragmentation. However, there is no consensus yet in the scientific community with respect to how to best analyse actigraphy data. Producing numerical indices condensing long minute-by-minute time-series into one single value may lead to loss of information. Here we describe using generalised additive models (GAMs) to analyse the entire time series. We apply GAMs to the analysis of minute-by-minute activity data in a real case study and show how GAMs can account for the different sources of variation specific to this type of data (e.g. time of the day, weekday vs. weekend) thus better exploiting the information contained in the actigraphy data. We also illustrate how GAMs can help to detect differences not only in volume of activity but also in the shape, i.e. in how the activity spreads over the day.

Munshi Imran Hossain
(Cytel)

Munshi Imran Hossain is an Associate Consultant in the Quantitative Strategies & Data Science group at Cytel. Imran has about 9 years of experience, working in the areas of software development and data science. He is a biomedical engineer by training and has interests in the processing and analysis of biomedical data including data from sensors, gene expression data and protein biomarker data among others. He is also a member of PSI's Special Interest Group for Data Science.

Processing and Analysis of Accelerometer Data in Subjects with Neurodegenerative Disorders.

Today, sensors have become an essential component of the healthcare ecosystem. They generate valuable data that can be processed to create insights to help improve healthcare. Wearable sensors take this a step further by allowing these systems to be "worn" by subjects. This allows for continuous monitoring and hopefully, timely alerts.

The challenge, however, is to process the vast stream of data generated from sensors and extract useful information. This is an involved exercise needing signal processing and statistical expertise. In this case study, I will present the processing and analysis of data generated from an accelerometer. The objective is to be able to detect difficulty in a physiological process among subjects with neurodegenerative disorders. I will give a glimpse into the kinds of features that can be extracted from raw accelerometer data which, in turn, can be used to build machine learning models.

Andrew Potter (CDER, FDA)

Dr. Andrew Potter is a mathematical statistician in the Division of Biometrics I in CDER supporting the review work in the Division of Psychiatry.  His research interests include the use of digital health technologies in clinical trials and the analysis of high frequency outcome data and in involved in working groups at FDA on this topic.  He received his PhD in Biostatistics from the University of Pittsburgh.

With recent developments in wearable biosensors and portable electronics (collectively referred to as digital health technology – DHT) , many clinical trials propose measuring endpoints using a DHT.  DHTs record data at a higher frequency than traditional clinician/patient observed data.  For example, a wearable accelerometer records data multiple times a second every day during follow-up (lasting multiple weeks).  Clinical endpoints are derived from this high frequency data.  Examples include:

• Trials where an experimental drug is hypothesized to improve exercise capacity with the endpoint of daily time spent at or above a specified exercise intensity
• Trials where daily sleep parameters are measured by a one or more sensors

To estimate and test treatment effects measured in these cases, several statistics issues must be addressed.  Unlike current endpoint measurements, DHTs provide endpoint measurements for every day worn.  In the case that different days (e.g., weekends vs. weekdays) are combined to form endpoints, treatment effect estimates may be biased or miss important disease features if days are not exchangeable.  With the greater amount of data collected, missing data can occur in multiple ways (e.g., a missing observation within a day, missing a day within a week, monotone dropout).

This talk discusses potential issues arising in exercise endpoints if a DHT is worn for different times each day during a study.  It also discusses a case study exploring how different statistical analyses of a clinical trial for a new insomnia drug are affected by missing data to motivate discussion of statistical considerations in studies using these new data collection tools.

Speaker

Biography

Abstract

Kirsty Hicks
(GSK)

Kirsty is a Senior Statistics Director at GSK, having joined in 2001 from Roche where she worked as a late phase statistician for a number of years. Throughout her time at GSK Kirsty has supported all gastrointestinal assets in the early phase portfolio covering diseases such as ulcerative colitis and irritable bowel syndrome; supported numerous immunological-inflammatory assets across a wide range of diseases; and more recently heads up the Cancer Epigenetics Biostatistics group. Throughout this time Kirsty has managed an ever growing team of statisticians in addition to being heavily involved in numerous non project initiatives. More recently these have included exploring experimental medicine analysis methods; facilitating prior elicitations and leading the biostatistics team investigating the incorporation and analysis of complex digital data into clinical trials. Kirsty also currently leads the UK Oncology Biostatistics team, working closely with Oncology Biostatistics teams in the US, India, Japan and China.

Advanced Analytics of Digital Data: A focus on sensor data.

What is the evolving area of Digital data you may ask? Nowadays most of us will be wearing a watch monitoring some aspect of our movement or use phone apps that collect health information. The data collected can be referred to as digital biomarkers, and this data, in particular sensor technology is becoming a common feature in clinical trials.

Digital data can be used to collect information on sleep patterns, respiration rate, step count and continuous monitoring of heart rate and energy expenditure. Collecting data through digital devices also increases patient engagement and can provide real time compliance monitoring. The regulations for use of digital technologies in clinical trials are still evolving, and current recommendations for the analysis of such data are limited.

Utilising sensors, and collecting actigraphy data, is a non-invasive method of monitoring activity. An actigraph sensor is worn for a time period (eg a week or more) to measure activity. The movements the actigraph undergoes are continually recorded; where 300 data points per second can be collected. As a statistician initial questions that spring to mind include ‘How can this volume of data be analysed or represented visually?’ Obvious summary measures would include average step count over a day; or the percentage of time a patient is active or sedentary. But a lot of information is lost, what about the raw data? For example, the energy used at the minute level or even second level? Statistical methods that can be applied to this data are under investigation and this presentation will introduce some of these.

Edoardo Lisi
(GSK)

Edoardo Lisi is a Clinical Statistician based in the UK. He has been at GSK since 2018, working in Phase 1 and Phase 2 studies. He is a member of GSK’s Advanced Analytics for Digital Data (AADD) group, where he is focused on the statistical analysis of physical activity data from wearable devices.

Using Generalised Additive Models (GAMs) to analyse wearable data.

Many diseases impact the ability of patients to perform physical activity. The use of digital wearable devices equipped with sensors that continuously collect actigraphy data in clinical studies represents a great opportunity to better understand patient populations and how diseases limit their mobility. However, wearable devices produce vast amounts of data (e.g. second-by-second or minute-by-minute over several weeks). The statistical analysis of such type of data to assess the impact of an intervention is challenging. Ideas proposed in the literature suggest to summarise physical activity through indices of total activity or indices of activity fragmentation. However, there is no consensus yet in the scientific community with respect to how to best analyse actigraphy data. Producing numerical indices condensing long minute-by-minute time-series into one single value may lead to loss of information. Here we describe using generalised additive models (GAMs) to analyse the entire time series. We apply GAMs to the analysis of minute-by-minute activity data in a real case study and show how GAMs can account for the different sources of variation specific to this type of data (e.g. time of the day, weekday vs. weekend) thus better exploiting the information contained in the actigraphy data. We also illustrate how GAMs can help to detect differences not only in volume of activity but also in the shape, i.e. in how the activity spreads over the day.

Munshi Imran Hossain
(Cytel)

Munshi Imran Hossain is an Associate Consultant in the Quantitative Strategies & Data Science group at Cytel. Imran has about 9 years of experience, working in the areas of software development and data science. He is a biomedical engineer by training and has interests in the processing and analysis of biomedical data including data from sensors, gene expression data and protein biomarker data among others. He is also a member of PSI's Special Interest Group for Data Science.

Processing and Analysis of Accelerometer Data in Subjects with Neurodegenerative Disorders.

Today, sensors have become an essential component of the healthcare ecosystem. They generate valuable data that can be processed to create insights to help improve healthcare. Wearable sensors take this a step further by allowing these systems to be "worn" by subjects. This allows for continuous monitoring and hopefully, timely alerts.

The challenge, however, is to process the vast stream of data generated from sensors and extract useful information. This is an involved exercise needing signal processing and statistical expertise. In this case study, I will present the processing and analysis of data generated from an accelerometer. The objective is to be able to detect difficulty in a physiological process among subjects with neurodegenerative disorders. I will give a glimpse into the kinds of features that can be extracted from raw accelerometer data which, in turn, can be used to build machine learning models.

Andrew Potter (CDER, FDA)

Dr. Andrew Potter is a mathematical statistician in the Division of Biometrics I in CDER supporting the review work in the Division of Psychiatry.  His research interests include the use of digital health technologies in clinical trials and the analysis of high frequency outcome data and in involved in working groups at FDA on this topic.  He received his PhD in Biostatistics from the University of Pittsburgh.

With recent developments in wearable biosensors and portable electronics (collectively referred to as digital health technology – DHT) , many clinical trials propose measuring endpoints using a DHT.  DHTs record data at a higher frequency than traditional clinician/patient observed data.  For example, a wearable accelerometer records data multiple times a second every day during follow-up (lasting multiple weeks).  Clinical endpoints are derived from this high frequency data.  Examples include:

• Trials where an experimental drug is hypothesized to improve exercise capacity with the endpoint of daily time spent at or above a specified exercise intensity
• Trials where daily sleep parameters are measured by a one or more sensors

To estimate and test treatment effects measured in these cases, several statistics issues must be addressed.  Unlike current endpoint measurements, DHTs provide endpoint measurements for every day worn.  In the case that different days (e.g., weekends vs. weekdays) are combined to form endpoints, treatment effect estimates may be biased or miss important disease features if days are not exchangeable.  With the greater amount of data collected, missing data can occur in multiple ways (e.g., a missing observation within a day, missing a day within a week, monotone dropout).

This talk discusses potential issues arising in exercise endpoints if a DHT is worn for different times each day during a study.  It also discusses a case study exploring how different statistical analyses of a clinical trial for a new insomnia drug are affected by missing data to motivate discussion of statistical considerations in studies using these new data collection tools.