Google AI Health Event: Everything Revealed in 13 Minutes

CNET
19 Mar 202412:39

Summary

TLDRスクリプトのエッセンスを提供する魅力的な要約で、ユーザーを引き込み、興味を引き起こす。

Takeaways

  • 🌟 Fitbit Labsを通じて、プレミアムユーザーが実験的なAI機能に早期アクセスし、フィードバックを提供するカスタムパーソナライズドなインサイトを得ることができます。
  • 🚀 Fitbit Labは、GoogleのAI専門知識を基に、マルチモダリティなタイムシリーズの健康データとフィットネスデータを統合し、意義のあるパーソナライズドなインサイトを提供するモバイルアプリ内の空間です。
  • 📊 データポイントの視覚化のためのグラフを生成することができ、例えば、より活発な日には睡眠スコアが良いことが発見されることがあります。
  • 📱 今後の Fitbit Lab機能は、モバイルアプリ内のFitbit Labsプログラムに登録されている限定的なAndroidユーザー数のためにテストのために利用可能になります。
  • 🤖 将来のAI機能を通じてパーソナライズされた健康体験を提供するために、Google Research Healthとウェルネス、専門医、パーソナライズされたトレーニングコーチと協力しています。
  • 🧠 個人健康LLMは、フィットネスからの高品質な研究ケーススタディに基づいて、健康とフィットネスのデータを理解し、トレーニングコーチのように調整された推奨を提供する、GIモデルのファインチューンされたバージョンです。
  • 🏥 Med LMという医学的に調整された大規模言語モデルを去年発表し、現在は分野で探求されていますが、世界中のパートナーがMedmと他のAIを革新的に使用する方法をいくつか強調します。
  • 🖋️ 米国のGeno Bioworksは、英国のHum Therapeuticsは医師をサポートし、インドのApollo病院は24/7のテレアヘルスサービスへのアクセスを容易にしています。
  • 📈 メタルアンドファミリーのモデルを拡張し、メタルX線胸部のメタルを始めとして、Google Cloudで実験的なプレビューで利用可能にしました。
  • 🤖 メタルの能力は、医療記録の分析や手順の簡素化など、医療従事者が最良のケアを提供するために役立ちます。
  • 🌐 検索で表示される自殺、家庭暴力、性暴力のホットラインを数十カ国言語に拡大し、今年はさらに20以上の国でカバーを拡大しています。
  • 🧠 AIの能力を評価し、世界中の人々にアクセスしやすいフォーマットで健康情報を提供することが重要です。

Q & A

  • Fitbit Labsとは何ですか?

    -Fitbit Labsはモバイルアプリ内のスペースで、プレミアムユーザーが実験的なAI機能に早期アクセスし、フィードバックを提供できるコンセプトです。

  • Fitbit Labsの目的は何ですか?

    -Fitbit Labsの目的は、ユーザーが新しいAI機能をテストし、フィードバックを通じてFitbitと協力しながらこれらの体験を構築し、改善することにあります。

  • Fitbit Labsで利用できる機能の例はありますか?

    -例えば、活動量が多い日に睡眠スコアが最も良いことを発見するなど、ユーザーの健康とウェルネスデータから有意義な洞察を導き出す機能があります。

  • 個人の健康データに基づいたカスタマイズされた推薦を提供するためにFitbitはどのようにGoogleのAI専門知識を活用していますか?

    -FitbitはGoogleのAI専門知識を活用して、ユーザーのマルチモーダルな時系列の健康とウェルネスデータを統合し、個々に合わせた洞察を導き出します。

  • Fitbitが開発している個人の健康に関する大規模言語モデル(LLM)の特徴は何ですか?

    -このモデルは、Fitbitのデータから得られた多様な健康シグナルに基づく高品質な研究ケーススタディを使用してファインチューニングされ、睡眠スケジュール、運動強度、心拍数変動などのパターンに基づいたよりカスタマイズされた洞察をユーザーに提供します。

  • FitbitとGoogleが健康とウェルネスの専門家と協力している理由は何ですか?

    -AIを活用して、より個人化された健康体験を提供し、ユーザーがフィットネス、健康、そしてウェルビーイングの目標を達成する新しい方法を開拓するためです。

  • Med LMとは何でしょうか?

    -Med LMは、医療分野で使用される、医学的に調整された大規模言語モデルで、医療ライセンス試験ベンチマークで合格スコアを達成した最初のAIシステムです。

  • FitbitとGoogleのAIが提供する健康情報を視覚的に理解しやすくする取り組みの例を教えてください。

    -Google Lensを使用して皮膚の写真を撮り、その画像に基づいてウェブから視覚的に類似したマッチを検索する機能があります。これにより、色や形、テクスチャを言葉で説明するのが難しい場合でも、健康情報を簡単に検索できます。

  • AIが医療分野で果たす役割についてFitbitとGoogleはどのようなビジョンを持っていますか?

    -FitbitとGoogleは、AIが医療診断の迅速化、医療提供プロセスの効率化、そして様々な健康問題に対するアクセスと理解の向上に重要な役割を果たすと考えています。

  • 健康情報のアクセシビリティを向上させるためにFitbitとGoogleはどのような取り組みをしていますか?

    -健康情報をよりアクセスしやすくするために、検索結果に自殺、家庭内暴力、性的暴行のホットラインを表示する国と言語を拡大し、抑うつや不安の臨床的に検証された自己評価をより多くの国で簡単に見つけられるようにしています。

Outlines

00:00

🤖 FitbitラボとAIエクスペリエンスのパーソナライズ

Fitbitラボは、プレミアムユーザーが実験的なAI機能に早期アクセスし、フィードバックを提供するモバイルアプリ内の空間です。GoogleのAI専門知識を活用し、フィットビットの多様な時間シリーズの健康データとウェルネスデータを統合し、チャートを生成することで、パーソナライズされた洞察を提供します。例えば、より活発な休日に最高の睡眠スコアが得られることを発見できます。今年後半には、Fitbitラボの機能が限定数のAndroidユーザーにテストとして利用可能になります。将来的には、Google研究の健康とウェルネス専門家、医師、個人トレーニングコーチと協力して、健康とフィットネスデータについて理論を立て、コーチのようにTAILORED RECOMMENDATIONを提供する個人健康LLM(Large Language Model)を開発する予定です。

05:03

📚 複雑なケースにおけるAIの医療記録の活用

電子医療記録には多くの患者情報がありますが、合成された医療記録をモデルの文脈の一部として提供することで、モデルは医師の質問に正確かつ直接的に答えることができます。AIの多様性を吸収し、テキスト、画像、オーディオなど異なるデータタイプを横断して理解する能力を持ち、医療専門家が最善のケア決定を行う際に信号を解釈するために必要な多様性を実現しました。3D脳CTの報告生成タスクにおいて、モデルが生成した報告の大部分は、独立した臨床医師によって手動で作成された報告と同等またはそれ以上の品質であると評価されました。しかし、AIシステムが放射線写真の報告を独立に生成する準備ができているかどうかを評価する基準がまだありません。

10:03

🌐 健康情報をより視覚的にアクセス可能に

健康情報をより視覚的にアクセス可能にするために、モバイルデバイスで視覚的な経験を向上させ、高品质なウェブからの画像や図を追加しました。これにより、頸痛などの症状を理解するのが容易になります。また、健康状態の視覚的な結果をモバイルで利用可能にするための作業も進めています。情報を受け取る最も理解しやすい形式であっても、テキスト、画像、動画を使用して、あなたの健康に関する質問に答えを見つける助けをします。次の数か月間で、このアップデートを世界中で展開する予定です。また、自殺、家庭内暴力、性的暴力のホットラインを検索に表示し、多くの国や言語にわたってカバーし、今年はさらに20以上の国でカバーを拡大します。これにより、人々は最も必要な時に地域のリソースに接続することができます。

Mindmap

Keywords

💡Fitbit Labs

Fitbit Labsは、フィットビットのモバイルアプリ内のエクスペリメンタルAI機能にプレミアムユーザーが早期アクセスできる空間です。この概念は昨年紹介され、ユーザーがフィードバックを提供することで、フィットビットと共同でAIエクスペリエンスを構築し、反復改善する機会を提供します。

💡AI experiences

AI experiencesは、人工知能技術を利用してユーザーに提供される新しいタイプの体験を指します。これらの体験は、健康やフィットネスデータの分析、視覚化、パーソナライズされた推奨など、多岐にわたる機能を含みます。

💡multimodal time series health data

多様なタイプのデータ(例えば心拍数、睡眠スコア、活動レベルなど)を時系列で分析し、健康状態やフィットネスレベルを理解するための方法です。

💡personalized insights

パーソナライズされた洞察は、個々のユーザーのデータに基づいて提供される分析結果や推奨を指します。これにより、ユーザーは自分に合った健康やフィットネスのアドバイスを受けることができます。

💡Google's AI expertise

GoogleのAI専門知識は、Googleが長年にわたって蓄積した人工知能技術や専門家知識を指します。これは、フィットビットラボのAIエクスペリエンスを構築する際に、高い品質と効率を保証するものです。

💡personal health llm

パーソナルヘルスLLM(Large Language Model)は、ユーザーの健康やフィットネスデータを理解し、パーソナライズされた推奨を行うために特別に調整された言語モデルです。

💡health and fitness goals

健康とフィットネスの目標は、個々人が健康を向上させ、体を鍛えるために設定する目標を指します。これには、体重減少、筋力向上、心肺機能の改善などが含まれます。

💡medically tuned large language model (Med LM)

メディカルリーチューンドのラージ言語モデル(Med LM)は、医療分野の知識やデータを使ってトレーニングされたAI言語モデルです。これにより、医療関連の質問や問題に対する正確な回答を提供できます。

💡multimodality

マルチモダリティは、複数のデータタイプ(テキスト、画像、音声など)を理解し、推理する能力を指します。これは、医療分野で特に重要で、医療従事者が最善の治療決定を行うために必要な能力です。

💡health equity

健康平等は、人種、年齢、社会経済的地位などの個人の属性に関係なく、すべての人が同じように高品質の医療サービスを受けられることを指します。

💡AI-powered search

AIパワード検索は、人工知能技術を利用して、ユーザーの検索クエリを解析し、より正確で関連性の高い結果を提供する検索方法です。

💡mental health crisis

精神的ヘルス危機は、個人が心の健康に関する深刻な問題や困難に直面し、その状況を管理できず支援が必要な状況を指します。

Highlights

Fitbit Labs is creating new AI experiences for personalized insights based on users' health and wellness data.

Premium users can get early access to experimental AI features in Fitbit Labs and provide feedback to help iterate on these experiences.

Fitbit Lab harnesses Google's AI expertise to analyze multimodal time series health data and generate personalized insights.

Users can visualize data points with charts, such as discovering sleep score patterns related to activity levels.

Fitbit Lab features will be available for testing on Android for a limited number of users enrolled in the program.

Fitbit is partnering with Google Research Health to create a personal health large language model (LLM) for tailored recommendations.

The personal health LLM is fine-tuned using high-quality, deidentified health data from Fitbit users.

The new personal health LLM will power future AI features across Fitbit's portfolio for personalized health experiences.

Med LM, a medically tuned large language model, is being explored in the field for various applications.

Geno Bioworks in the US is using AI to advance drug discovery and biocurity.

Hum Therapeutics in the UK supports clinicians with better insights for patient care.

Apollo hospitals in India use AI to ease access to their 24/7 telehealth services.

Metaland family of models is expanding to include multimodality, starting with metm for chest x-ray on Google Cloud.

Medam for chest x-ray enables findings classification, semantic search, and more to improve radiologist workflows.

AI systems like Med LM can assist in the nurse-to-nurse handoff process by creating comprehensive summaries.

AI models are being fine-tuned for the medical domain, achieving state-of-the-art performance on medical QA benchmarks.

AI can provide synthetic medical records to help physicians answer questions about patient histories more efficiently.

There's a lack of representation of marginalized populations in clinical trial research, which needs to be corrected for equitable AI model development.

Heel (Health Equity machine learning assessment) is used to evaluate AI models for equitable performance across different demographics.

AI models are being built to diagnose diseases sooner and provide life-saving treatment, with Apollo working to incorporate these models into national screening programs in India.

Google Lens allows users to search for health information by taking a picture of a skin condition, providing visually similar matches from the web.

Health information on mobile devices is being made more visual to improve understanding of symptoms and health conditions.

Search engine coverage for suicide, domestic violence, and sexual assault hotlines is being expanded to more countries and languages.

Clinically validated self-assessments for mental health are being made easier to find in more countries.

YouTube's AI-powered tool streamlines video translation and dubbing, empowering creators to reach wider audiences.

Language access is crucial for providing quality health information, exemplified by free Spanish courses on promoting racial justice in medical education.

Transcripts

play00:00

to bring tailored personalized insights

play00:03

based on your unique needs and

play00:05

preferences that's why we're creating

play00:08

new AI experiences in Fitbit Labs it's a

play00:11

concept that we introduced last year

play00:14

Fitbit lab would be a space in the

play00:16

mobile app where premium users can get

play00:19

early access to experimental AI features

play00:22

so that they can test out and give

play00:25

feedback this offers an opportunity for

play00:27

our users to partner with us as we build

play00:30

and iterate on these experiences

play00:32

building on Google's AI expertise Fitbit

play00:36

lab can help you derive meaningful and

play00:38

personalized Insight by bringing

play00:40

together your

play00:42

multimodal time series health and

play00:44

wellness data it can even generate

play00:46

charts for the data points that you want

play00:49

to visualize for example you could

play00:51

discover that your sleep score is best

play00:54

on the days that you are more active on

play00:56

a recent vacation with my family I was

play00:59

put in charge of all the children one

play01:00

day and I hit my record number of steps

play01:03

and then my best sleep scored that night

play01:06

with these type of tools and features

play01:08

I'll be able to then dig deeper into

play01:10

these type of connection am I just more

play01:12

active when I'm around children um later

play01:15

this year Fitbit lab features will be

play01:17

available for testing for a limited

play01:19

number of Android users who are enrolled

play01:22

in the Fitbit Labs program in the mobile

play01:25

app looking ahead we want to deliver

play01:29

even more personalized Health

play01:31

experiences with AI so we're partnering

play01:34

with Google research Health and Wellness

play01:36

expert doctors and personalized and

play01:40

certified coaches to create personal

play01:43

health large language model that can

play01:46

reason about your health and fitness

play01:49

data and provide tailored recommendation

play01:52

similar to how a personal coach would

play01:55

this personal health llm will be a

play01:58

fine-tune version of our G I model this

play02:01

model fine-tuned using high quality

play02:04

research case studies based on the

play02:06

deidentified

play02:08

Diverse Health signals from Fitbit will

play02:11

help users receive more tailored

play02:13

insights based on patterns in sleep

play02:16

schedule exercise intensity changes in

play02:19

heart rate variability resting heart

play02:21

rate and many many more this new

play02:23

personal health llm will power future AI

play02:26

features across our portfolio to bring

play02:30

personalized Health experiences to our

play02:33

users while it's not meant to diagnose

play02:37

treat mitigate cure or prevent any

play02:40

disease injury or

play02:43

condition we hope that this more

play02:45

personalized AI coaching model can help

play02:47

you reach your fitness health and

play02:50

well-being goals in ways that were not

play02:52

possible before just last year at the

play02:54

checkup we launched our medically tuned

play02:57

large language model Med LM and fast

play03:00

forward to today it's already being

play03:02

explored in the field you'll hear

play03:04

several examples but I want to highlight

play03:07

three ways our partners around the world

play03:09

are using medm and our other AI to

play03:13

innovate in the US Geno bioworks is

play03:16

advancing drug Discovery and

play03:18

biocurity in the UK hum Therapeutics is

play03:21

supporting clinicians with better

play03:24

insights and in India Apollo hospitals

play03:27

is using AI to ease access to their 24/7

play03:31

teleah Health Services we're now

play03:33

expanding our metaland family of models

play03:35

to include multimodality starting with

play03:38

metm for chest x-ray which is now

play03:41

available in an experimental preview on

play03:43

Google Cloud to allow our customers to

play03:45

build solutions medam for chest x-ray

play03:48

will enable things like findings

play03:49

classification semantic search and more

play03:53

we hope this will provide solutions that

play03:55

can improve the efficiency of

play03:57

radiologist workflows empowering them to

play03:59

the Del high quality and consistent care

play04:02

when we first showed medm to members of

play04:04

our care team we heard a resounding

play04:06

feedback that this could be extremely

play04:08

valuable in supporting part of the Care

play04:10

process that has been an issue since the

play04:12

beginning of healthcare the nurse to

play04:14

nurse handoff sometimes commonly known

play04:16

as the bedside shift report Metal's

play04:18

capabilities can be used to analyze

play04:21

Patients health history clinician notes

play04:23

and more to create a comprehensive easy

play04:26

to use summary for the handoff a nurse

play04:30

then reviews the note confirms its

play04:33

accuracy modifies as needed and hands it

play04:36

over to the counterart this helps nurses

play04:39

quickly get to the information they need

play04:42

about a year and a half ago we built the

play04:44

first AI system that was able to achieve

play04:47

a passing score on a medical licensing

play04:50

exam Benchmark called

play04:51

mqa our newest model fine-tuned for the

play04:55

medical domain achieves state-of-the-art

play04:57

performance on medical on the medical QA

play04:59

Benchmark standing at over 91% when

play05:02

doctors see new patients particularly in

play05:04

complicated cases they have questions

play05:07

about the patient medical history much

play05:10

of this information is held in dense

play05:12

electronic medical records but now we've

play05:15

observed that if we provide synthetic

play05:18

medical record a synthetic medical

play05:20

record as part of the context our models

play05:22

are able to correctly and directly

play05:25

answer a physician's questions about the

play05:28

patient's history multimodal it is the

play05:30

ability to absorb and reason across

play05:34

multiple different data types like text

play05:37

images and

play05:39

audio Gemini models were built from the

play05:42

ground up to be

play05:45

multimodo and after all medicine is also

play05:49

inherently multimodal to make the best

play05:53

care decisions Healthcare professionals

play05:56

regularly interpret signals across a

play05:58

plethora of source

play06:00

including medical images clinical notes

play06:03

electronic health records and lab tests

play06:05

we tested how our new model would

play06:06

perform on the significantly more

play06:08

complex task report generation for 3D

play06:11

brain

play06:13

CTS with these 3D images we found that a

play06:16

significant portion of the reports that

play06:18

our model generated were judged by

play06:20

independent clinicians to be on a par or

play06:23

better than manually created

play06:26

reports well these results are very

play06:29

encouraging ing they're not an

play06:31

endorsement that these AI systems are

play06:33

ready to be trusted to generate

play06:35

Radiology reports

play06:37

independently instead they emphasize

play06:40

that it's time to evaluate ai's ability

play06:43

to assist Radiologists in Real World

play06:46

Report generation workflows there's

play06:48

currently no established standard to

play06:50

ensure that data used to develop AI

play06:53

reflects the diversity of people and

play06:55

experiences around the world and in

play06:58

healthcare there's historically been a

play07:01

lack of representation of historically

play07:04

marginalized populations in areas like

play07:07

clinical trial research we realized that

play07:10

many existing Dermatology data sets

play07:12

include primarily skin cancers and other

play07:15

severe conditions yet lack common

play07:18

concerns like allergic reactions it's

play07:20

estimated that less than 20% of

play07:23

Dermatology textbook images contain dark

play07:27

skin tones a statistic that has has not

play07:30

changed in about 15 years to correct the

play07:34

eror the failures of the past we need to

play07:37

ensure these biases are not repeated in

play07:41

in the way that we build AI models we

play07:44

are using heel which we've determined is

play07:47

called Health Equity um assessment of

play07:51

machine learning and we dubbed it heel

play07:53

for short using heel we evaluated an AI

play07:57

model designed to predict Dermatology

play08:00

conditions based on photos of skin

play08:03

concerns the model performed equitably

play08:06

across race ethnicity and skin subgroups

play08:11

but we discovered that it had some room

play08:13

for improvement when it came to age

play08:16

older adults 70 and over are at risk of

play08:21

worse Health

play08:22

outcomes from skin conditions our model

play08:26

recognized that with Cancers but it did

play08:29

not when it came to more common concerns

play08:32

like allergic

play08:33

reactions this is a great example of the

play08:36

heel framework doing what it was meant

play08:39

to do highlighting where we need to

play08:42

improve our model today we're excited to

play08:44

share that aollo will build Upon Our TB

play08:47

lung and breast cancer models to help

play08:50

diagnose more people sooner we believe

play08:53

that AI has an important role here in

play08:56

helping more people receive diagnosis

play08:58

sooner and get life-saving

play09:01

treatment Apollo is working towards

play09:04

bringing these models to markets across

play09:06

India this means that with the required

play09:08

regulatory approvals they can

play09:10

incorporate our algorithms into

play09:12

screening programs

play09:14

nationally additionally over the next 10

play09:17

years Apollo will provide AI power

play09:19

screenings for TB lung and breast cancer

play09:22

in under resource communities at no cost

play09:25

some health questions are really

play09:28

difficult to describe in words alone for

play09:31

example if you see a discoloration or

play09:33

other abnormality on your skin trying to

play09:36

describe the color or the shape or the

play09:37

texture is not always easy so starting

play09:41

last year we made it possible to use

play09:44

Google Lens to search what you see on

play09:46

your skin and it couldn't be easier take

play09:49

a picture of your skin with lens in the

play09:51

Google app and you'll find visually

play09:54

similar matches from the web to inform

play09:56

your search this AI powered feature

play09:59

available in more than 150 countries

play10:02

also works if you're not sure how to

play10:04

describe something else on your body

play10:06

like a bump on your lip another part of

play10:08

making health information accessible to

play10:10

everyone is presenting it in formats

play10:12

they can easily understand so with this

play10:16

in mind we've been making the experience

play10:17

more visual on mobile devices we've

play10:20

added images and diagrams from high

play10:22

quality sources on the web that make it

play10:24

easier to understand symptoms like neck

play10:27

pain for example

play10:29

and we're also working to make these

play10:31

more visual results available on mobile

play10:33

for health conditions as well such as

play10:35

migraines kidney stones and pneumonia in

play10:38

whichever format you understand

play10:41

information best using text images or

play10:43

videos we want you to help want to help

play10:46

you find those answers to your health

play10:48

questions and over the next few months

play10:50

we'll be rolling out this update

play10:51

globally we've extand expanded the

play10:53

suicide domestic violence and sexual

play10:55

assault hotlines shown in search to

play10:57

dozens of countries and languages this

play11:01

year alone we'll increase coverage

play11:02

across these features by 20 additional

play11:05

countries including Puerto Rico and

play11:07

Thailand this will help people connect

play11:10

with local resources when they need it

play11:13

most and as searches for Mental Health

play11:16

crisis continue to climb year after year

play11:19

we've made it easier to find clinically

play11:22

validated self assessments for

play11:24

depression and anxiety in more countries

play11:26

YouTube's no cost AI power tool

play11:30

streamlines the video translation and

play11:32

dubbing process empowering creators to

play11:34

expand their reach and reaching people

play11:36

where they are begin two compressions

play11:39

per second or 100 to 120 per minute and

play11:43

press down at least two Ines in depth

play11:46

with the heel of your

play11:47

[Music]

play11:56

hand imagine a Spanish speaking mother

play12:00

learning how to perform CPR on her child

play12:03

a construction worker studying how to

play12:05

control bleeding in case of an accident

play12:08

even thinking through the cases of an

play12:09

elderly couple being able to quickly

play12:12

identify signs of a stroke and these are

play12:14

just a couple of examples of how

play12:17

important it is to really break down

play12:18

this issue of language access and give

play12:20

people quality information when they

play12:22

needed the most starting today an

play12:24

animation style course on how to promote

play12:26

racial Justice in medical education is

play12:29

available in Spanish for free on the

play12:32

Stanford medicine continuing medical

play12:33

education Channel thanks to allow and

play12:35

shout out to the people from Stanford in

play12:37

terms of their leadership in this as

play12:38

well

Rate This

5.0 / 5 (0 votes)

Related Tags
フィットビットAI体験パーソナライズ健康データGoogleラボフィードバック実験的モバイルアプリ
Do you need a summary in English?