COLLEGE OF BUSINESS AND ECONOMICS

Department of Computer Science

Course: Machine Learning

Revolutionizing Healthcare with AI: Comparing DeepMind

Health, IBM Watson Health, and Tempus
GROUP Assignment
Names ID no
1. Bosona Gosaye……………………………..…UGR/12161/14
2. Mekuanint Liyew…………………………..…UGR/10696/13
3. Magist Haile……………………………..……UGR/11279/14
4. Saad Zenu………………….……………….…UGR/10838/13
5. Mohammed Abdurhaman……………………..UGR/11092/13

I
Table of Contents
Revolutionizing Healthcare with AI: Comparing DeepMind Health, IBM
Watson Health, and Tempus I
Table of Contents II
Abbreviation II
1. Introduction 1
2. Background and Theory of AI in Healthcare 1
3. Related Work in AI Healthcare 2
4. DeepMind Health 2
Clinical Applications and Outcomes 2
5. IBM Watson Health 3
Technical Architecture and Algorithms 3
Clinical Applications and Outcomes 3
6. Tempus 4
Data and AI Techniques 4
Clinical Applications and Products 5
Collaborations and Outcomes 5
7. Discussion and Analysis 6
Strengths and weaknesses: 6
8. Conclusion and Future Work 7
References 8

II
Abbreviation
· AI – Artificial Intelligence

· ML – Machine Learning

· DL – Deep Learning

· NLP – Natural Language Processing

· EHR – Electronic Health Record

· CAD – Computer-Aided Detection

· AMD – Age-Related Macular Degeneration

· OCT – Optical Coherence Tomography

· CoDoC – Complementarity-Driven Deferral-to-Clinical Workflow

· AKI – Acute Kidney Injury

· NGS – Next-Generation Sequencing

· RNA – Ribonucleic Acid

· DNA – Deoxyribonucleic Acid

· xT – Tempus Tumor DNA/RNA Panel (brand name)

· nP – Tempus Neuropsychiatric Panel (brand name)

· ECG – Electrocardiogram

· ECG-AF – Tempus ECG-based Atrial Fibrillation risk model

· FDA – Food and Drug Administration

III
IV
1. Introduction
Artificial intelligence (AI) is rapidly proliferating across sectors, and healthcare has
emerged as a primary arena for its transformative potential . By processing large
volumes of medical data, AI promises to revolutionize patient care and streamline
operations. In medicine, machine learning (ML) and deep learning techniques excel at
pattern recognition tasks such as medical imaging analysis and genomics
interpretation . For example, deep neural networks have demonstrated accuracy in
cancer detection that can rival or exceed previous computer-aided tools . Similarly,
natural language processing (NLP) algorithms can parse clinical notes and literature,
enabling intelligent report generation and decision support Together, these
capabilities suggest AI could greatly improve diagnostics, risk stratification, and
treatment planning in clinical settings .

2. Background and Theory of AI in Healthcare

AI in healthcare encompasses a variety of methods for analyzing complex medical
data. At its core, machine learning (ML) refers to algorithms that learn patterns from
data. This includes supervised learning (training on labeled examples) and
unsupervised learning (finding structure without explicit labels). Deep learning, a
subfield of ML, uses multi-layer neural networks (often with millions of parameters)
to recognize intricate patterns. Modern deep learning benefits from powerful GPUs
and large datasets, making it suitable for tasks like medical image interpretation. In
radiology, for example, convolutional neural networks have been applied to detect
cancers on imaging with high sensitivity . These deep networks can learn from raw
pixels or radiomic features to surpass older computer-aided detection (CAD)
methods .
Another important AI domain is Natural Language Processing (NLP). NLP enables
computers to interpret and generate human language. In healthcare, NLP systems
analyze clinical notes, pathology reports, and literature. They can extract key findings
from unstructured text, summarize patient records, or generate draft reports. For
instance, machine learning–based NLP algorithms can identify diagnoses and
medications from doctors’ notes . There are two main approaches: statistical (often
deep learning models trained on large text corpora) and knowledge-based. In practice,
many healthcare AI platforms use NLP to process free-text data and integrate it with
structured data for decision support.
Historically, some AI systems in medicine were rule-based expert systems, where
human experts encoded if-then rules. These were popular in earlier decades for
clinical decision support, but they require manual knowledge engineering and do not
scale easily . Today, such systems are largely supplanted by data-driven ML
approaches. However, expert systems taught us the importance of clinical context and
domain knowledge. Modern AI efforts often blend statistical learning with curated
knowledge bases for improved reasoning.
Large-scale data and computational advances drive the recent AI surge. Healthcare
generates “big data” in volume and complexity – from EHRs to genomics. AI
algorithms can leverage these vast datasets to learn complex patterns that elude
traditional methods . AI in healthcare is not a single technology but a toolkit –
including deep neural networks, ML classifiers, NLP pipelines, and even
reinforcement learning – all deployed to improve diagnosis, treatment, and workflow
efficiency.

1
3. Related Work in AI Healthcare
Beyond the three systems at hand, many organizations are contributing AI innovations
in healthcare. In medical imaging, companies like NVIDIA (Clara) and Siemens
Healthineers are developing deep learning tools for radiology and MRI analysis.
Startups such as PathAI and PaigeAI apply deep learning to pathology slides for
cancer detection. Tech giants like Microsoft have projects (e.g. Project InnerEye)
focusing on image segmentation, while Amazon (AWS HealthLake) and Apple
(Health Records) invest in health data infrastructure. In genomics, firms like
Illumina, Flatiron Health (cancer data), and Adaptive Biotechnologies combine AI
with genetic testing and real-world evidence.
A notable research breakthrough is AlphaFold from DeepMind, which uses deep
learning to predict protein structures. In 2020, AlphaFold accurately modeled the 3D
structure of complex proteins like CDK20, enabling rapid drug design (a candidate
liver cancer inhibitor was developed in just 30 days) . This exemplifies how AI is
extending beyond diagnostics into drug discovery. Similarly, AI-driven drug
discovery startups (e.g. Insilico Medicine, Atomwise) use deep learning to screen
molecules.
Government and academia also push AI. For example, DARPA and NIH fund
initiatives to use ML for population health and diagnostics. Many healthcare AI
survey studies and frameworks have emerged, emphasizing human-centered design
and iterative validation. These related efforts underscore a vibrant ecosystem where
DeepMind Health, IBM Watson, and Tempus are only part of a larger wave
transforming medicine with AI.

4. DeepMind Health
DeepMind Health (now part of Google Health) applies cutting-edge AI research to
clinical problems. Founded as part of Google in 2016, DeepMind collaborated with
hospitals (e.g. Moorfields Eye Hospital in London) to develop diagnostic algorithms.
A hallmark project is the Moorfields Age-Related Macular Degeneration (AMD)
study. Researchers collected optical coherence tomography (OCT) retinal scans from
thousands of patients and trained a two-stage deep convolutional neural network. The
first network segmented anatomical structures; the second combined raw and
segmented data to predict conversion from dry to wet (exudative) AMD within six
months. The system achieved performance on par with retinal specialists,
demonstrating that AI can effectively predict disease progression.
DeepMind’s technical approach emphasizes deep learning and hybrid models. In the
AMD model, inputs included 58 million 3D OCT pixels, illustrating the scale of data
used. The model architecture fuses image segmentation and raw-image analysis,
enabling it to learn both known clinical features (like drusen) and subtle patterns.
Another project, CoDoC (Complementarity-Driven Deferral-to-Clinical Workflow), is
detailed in Nature Medicine (2023) . CoDoC trains a meta-model to decide when to
trust an AI prediction versus defer to the clinician. In simulated breast cancer
screening, CoDoC reduced false positives by ~25% without missing true cancers .
This highlights DeepMind’s focus not just on AI accuracy but also on safe integration
with clinicians.
Clinical Applications and Outcomes

2
DeepMind Health’s AI tools have targeted several areas:
Ophthalmology (AMD): As noted, the retinal AI predicted wet AMD onset
as well as human experts , suggesting utility for early intervention.
Mammography: CoDoC addresses mammogram interpretation by learning
when AI or radiologists are each more reliable. Although still experimental,
this human-AI collaboration model reduced unnecessary workups in testing.
Sepsis and Kidney Alerts: DeepMind’s earlier “Streams” app (2016) aimed
to detect acute kidney injury (AKI) by sending alerts to clinicians. While
Streams showed technical promise (identifying at-risk patients), it faced
controversy over NHS data governance and was transitioned out of
DeepMind. This episode underscores ethical/data concerns in real-world AI
deployments.
Protein Folding (AlphaFold): While not part of “DeepMind Health”,
AlphaFold (by DeepMind) has major healthcare implications. In early 2020,
AlphaFold’s protein structure predictions enabled rapid development of
potential cancer therapies. This achievement demonstrates DeepMind’s
strength in fundamental AI research with downstream medical impact.
DeepMind Health excels in research-driven AI with high algorithmic performance. Its
innovations (e.g. CoDoC) address technical challenges of reliability. However, real-
world clinical adoption has been limited; besides pilot studies, most tools remain
research prototypes. Privacy and regulatory hurdles (as seen in the Streams NHS case)
present obstacles. DeepMind’s advantages lie in cutting-edge models and strong
research teams, but it lacks an established product line and healthcare market presence
compared to other vendors.

5. IBM Watson Health

IBM Watson Health emerged from IBM’s AI ambitions. Watson famously won
Jeopardy! in 2011 by mastering natural language questions. Building on this, IBM
launched Watson Health in 2015, aiming to apply “cognitive computing” to
medicine . IBM acquired several companies (e.g. Merge Healthcare in 2016 , Truven
Health Analytics in 2016 to build a vast health data repository (~300 million patient
records . The goal was to feed Watson algorithms with rich clinical data. Watson
Health’s platform included Watson for Oncology (trained with Memorial Sloan
Kettering protocols) and Watson for Genomics, among others.
Technical Architecture and Algorithms
Watson Health leveraged NLP and rule-based analytics rather than purely deep
learning. Its core engine ingests patient records, medical literature, and guidelines,
applying NLP to parse symptoms, genomics, and medical terms. For example,
Watson for Oncology analyzed a patient’s tumor genetics and history, cross-
referenced oncology databases, and suggested treatment options. Internally, Watson
used a hybrid of statistical NLP, semantic analysis, and a knowledge graph. It also
employed machine learning models on structured data (e.g. lab values) and many
expert-curated rules. Watson’s architecture was not transparent; it often acted as a
“black box” advisor, which later drew criticism .
Clinical Applications and Outcomes
IBM marketed Watson Health for several applications:
Cancer care: Watson for Oncology was deployed in many hospitals
worldwide. IBM claimed Watson could analyze a brain cancer genome in ~10

3
 minutes (a task taking humans ~160 hours) . However, independent reports
found that early versions produced unsafe or incorrect treatment
recommendations. For instance, one STAT News investigation (2018) noted
multiple “unsafe and incorrect” suggestions in oncology trials. Internal
reviews revealed that some of Watson’s advice was based on hypothetical
rather than real patient data. These issues underscored that Watson’s AI
sometimes struggled to generalize.
Genomics: Watson for Genomics aimed to interpret genetic variants. It
matched mutations to known therapies by scanning research literature. While
fast, its clinical adoption was limited. Some health systems replaced Watson
genomics with more specialized tools or local expertise.
Imaging: After acquiring Merge Healthcare, IBM worked on Watson
Imaging, applying AI to radiology workflows. Merge brought advanced image
management technology, and IBM intended to overlay analytics on these
images . However, no major FDA-cleared “Watson Radiology” tool emerged
publicly.
Operational/Research tools: Watson Health also included tools like Watson
Health Insights (analytics for payers and pharma). These used ML to predict
outcomes or optimize clinical trials, but they were more data-analytics
platforms than clinical AI.
Despite the hype, empirical success was limited. Studies on Watson for Oncology
showed low concordance with tumor board decisions. Many pilot projects did not
translate to measurable patient outcomes. By 2020, revenues declined and IBM scaled
back Watson Health investments. In January 2022 IBM announced sale of Watson
Health’s data and analytics assets (e.g., MarketScan claims data, Micromedex) to
Francisco Partners, focusing IBM back on cloud/AI core. Watson Health’s
commercial trajectory serves as a caution: technological prowess alone does not
guarantee clinical impact. Issues included inconsistent recommendations, lack of
model interpretability, and difficulties integrating into clinical workflows.
IBM Watson Health pursued a broad “cognitive” approach: unstructured data (text,
images), ML, and an expansive knowledge base. It amassed one of the industry’s
largest healthcare datasets through acquisitions. However, its outcomes fell short of
promises. Watson’s strength in NLP and enterprise support was undermined by
reliability issues in practice. Its primary weakness was the gap between AI
suggestions and clinician judgment – Watson often required significant local
customization. In summary, Watson Health demonstrated the challenges of applying
AI to complex medical decision-making, underscoring that accuracy and validation
are crucial.

6. Tempus
Tempus is a Chicago-based precision medicine company founded in 2015 by
entrepreneur Eric Lefkofsky. Its business model centers on combining large-scale
clinical/genomic data with AI to guide oncology care. Tempus has built what it calls
“one of the world’s largest libraries of clinical and molecular data” reportedly
millions of patient records with paired genomic profiles and outcomes. The
company’s “operating system” includes data infrastructure and ML tools for analyses.
Data and AI Techniques

4
Tempus focuses on multimodal data integration. Its inputs include next-generation
sequencing (NGS) of tumors, RNA expression, pathology images, lab results, and
EHR data. Using these, Tempus develops ML models to identify biomarkers and
suggest personalized therapies. Unlike the pure DL focus of DeepMind, Tempus
relies on a mix of statistical models and machine learning pipelines that incorporate
genomic features. Notably, Tempus is embracing the new trend of large AI foundation
models. In April 2025 Tempus, along with AstraZeneca and Pathos AI, announced a
$200M collaboration to build an oncology “multimodal foundation model” trained on
Tempus’s vast oncology dataset. This indicates Tempus’s strategic pivot toward
generative AI and transfer learning in medicine.
Clinical Applications and Products
Tempus’s flagship offerings are its genomic testing services:
xT (Oncology Panel): A targeted DNA panel for solid tumors, sequencing
~600 cancer-related genes to identify mutations and resistance mechanisms.
xT results inform oncologists on targeted therapies and clinical trial eligibility.
A Nature Biotechnology study (2020) found Tempus’s xT test increased
opportunities for personalized therapy by combining DNA and RNA
profiling .
nP (Neuropsychiatric Panel): A pharmacogenomic test covering dozens of
psychiatric medications. It predicts likely efficacy and side effects, aiming to
personalize mental health treatment. (The company has case reports on its
efficacy.)
Tempus ECG-AF: An AI algorithm that analyzes 12-lead ECGs to predict
risk of future atrial fibrillation. In late 2023, Tempus received FDA 510(k)
clearance for ECG-AF, the first ML-based AF risk detection tool.
Olivia App: In 2024 Tempus beta-launched "Olivia," an AI-powered personal
health concierge app . It lets patients aggregate their health records and ask
questions via generative AI. This reflects Tempus’s push into patient-facing
AI tools.
Collaborations and Outcomes
Tempus has strategically partnered with pharma and institutions:
It began a multi-year collaboration with GSK (2022) to use Tempus’s AI
platform and data to improve drug discovery and trial design (.
In 2025, Tempus announced the AstraZeneca/Pathos multimodal AI project
mentioned above.
It also collaborates with BioNTech, leveraging Tempus’s datasets and
computational biology for immuno-oncology R&D.
As of 2024, Tempus’s partnerships include 95% of top pharma oncology
companies and over half of US oncologists, illustrating broad industry
adoption.
On the business side, Tempus reported robust growth: 33% revenue increase in Q3
2024 compared to prior year. Its core strength is the integration of clinical practice
with data analytics. By sequencing tumor samples in real clinical settings, Tempus
continuously enriches its dataset and validates AI findings. For example, analyses of
Tempus data have identified new drug targets and trial matches. The company also
provides "Data and Services" to payers and biopharma, applying ML to improve trial
enrollment and outcome prediction.
Tempus’s approach is data-driven precision medicine. It leverages large curated
datasets and focuses on actionable outputs for oncology care . Compared to

5
DeepMind’s research focus and Watson’s broad cognitive aim, Tempus occupies a
middle ground: it is a commercial entity applying AI pragmatically. Its strengths
include its enormous data repository and close ties to clinical workflows. This allows
Tempus to achieve tangible gains (e.g., more patients identified for trials, as in their
press kit. Its weaknesses include a narrow focus (primarily oncology and some
specialty areas) and the challenge of integrating rapidly evolving AI models into
regulated healthcare. Tempus’s pivot to generative AI models signals a strategy to
stay at the forefront of AI, building on its decade of data collection to create next-gen
clinical tools.

7. Discussion and Analysis

Comparing these three systems reveals complementary strengths and common
challenges.
DeepMind Health brings cutting-edge AI algorithms (e.g. deep CNN architectures)
and strong performance in research settings . Its emphasis on novel techniques (like
CoDoC) addresses issues of safety and collaboration. However, DeepMind lacks a
commercial product; its tools remain primarily at the pilot stage. There have also been
privacy and trust concerns (e.g., the NHS AKI app controversy).
IBM Watson Health took a very different approach: ingesting massive datasets and
applying NLP-based “cognitive” analysis. It had the advantage of scale (e.g. ~300M
patient records and branding. Yet in practice, Watson systems often produced
inconsistent results . The overreliance on proprietary models and limited real-world
validation were key weaknesses. By 2025 Watson Health no longer operates as an
IBM priority, illustrating the difficulty of delivering AI in healthcare at scale.
Tempus is essentially a data company with ML. Its strength lies in pragmatism:
focusing on specific use cases (oncology, cardiology risk) and building evidence of
impact. For instance, Tempus’s FDA-cleared ECG-AF algorithm and its expanding
genomic test menu show direct clinical applications. Tempus also forges industry
partnerships to apply its data to drug development, a niche neither DeepMind nor
Watson pursued as vigorously. However, Tempus depends heavily on genomic data –
its models may not generalize outside oncology or the patient populations it serves. It
also faces challenges common to all AI in health: ensuring model fairness, obtaining
regulatory approval, and convincing clinicians to trust AI recommendations.
All three systems grapple with AI’s “black box” problem. As one review notes, deep
models often lack interpretability, making it hard to audit errors. This appeared in
Watson’s unsafe cancer suggestions and remains a concern for Tempus’s predictive
models. On the other hand, DeepMind’s CoDoC explicitly tackles this by deferring
ambiguous cases to clinicians , an example of hybrid human-AI design that could be
emulated more widely.
key findings include:
 DeepMind excels in AI innovation (e.g. retina analysis accuracy, human-AI
workflows but has limited deployment.
 Watson Health had the vision of large-scale cognitive AI but struggled with
clinical reliability and ultimately wound down.
 Tempus has translated AI into clinical tools (genomic tests, risk scores) at scale,
leveraging its data assets, but its impact is mostly in oncology and related fields.

Strengths and weaknesses:

 DeepMind’s strengths are technical prowess and validation against specialists;
weaknesses are deployment and regulatory acceptance.

6
 Watson’s strength was data breadth and enterprise integration; weakness was
accuracy and clinical outcomes .
 Tempus’s strengths are vast data and clear clinical use cases; weaknesses include
narrow focus and reliance on continual funding (it has not yet been profitable).
Common limitations include data quality (hospital data can be messy), patient
privacy considerations, and the need for clinical trials to prove benefit. Notably, each
system has faced criticisms related to data governance or AI “explainability.”

8. Conclusion and Future Work

AI is reshaping healthcare, but its promise comes with caveats. This review compared
three pioneers: DeepMind Health, IBM Watson Health, and Tempus. We found that
DeepMind’s algorithms demonstrate that AI can match clinicians in tasks like retinal
disease prediction, yet real-world implementation remains challenging. Watson
Health showed that assembling large data and applying NLP is a powerful strategy,
but practical utility was limited and eventually stalled. Tempus exemplifies a data-
centric approach: by amassing clinical/genomic records and developing targeted AI
tools, it has created marketable precision medicine products (with industry
collaborations) .
Each system has strengths to build on: DeepMind’s research leadership, Watson’s
data infrastructure, and Tempus’s clinical integration. Yet weaknesses are evident:
DeepMind must focus on clinical partnerships and ethical data practices; IBM’s
Watson model highlighted the need for transparency and robust validation; Tempus
must diversify beyond oncology and prove its AI models improve patient outcomes.
Future research should emphasize multimodal integration and human-AI
teaming. The trend of developing large AI models is clear Tempus and partners are
pioneering multimodal foundational models for oncology. Similar efforts should
extend to other diseases. Enhancing interpretability is also crucial: techniques like
CoDoC’s deferral mechanism show promise in making AI safer. Clinician-in-the-
loop systems and rigorous prospective trials will be needed to establish trust.
A major suggestion is to create standardized evaluation frameworks for medical
AI. Rigorous benchmarks, shared datasets, and external validations (beyond vendors’
internal tests) would help distinguish genuine advances from hype. Regulatory bodies
(e.g. FDA, EMA) are already crafting AI-specific guidelines; future work must align
with these to facilitate approvals. Collaboration between tech companies, healthcare
providers, and regulators is key.
AI is indeed revolutionizing healthcare, but success requires not only advanced
algorithms but also careful integration into clinical practice. DeepMind Health, IBM
Watson, and Tempus each illustrate different facets of this journey – from lab to
patient. By learning from their experiences, future AI efforts can better harness
technology to deliver safer, more effective, and more equitable healthcare solutions.

7
References
[1] M. Chustecki, “Benefits and Risks of AI in Health Care: Narrative Review,”
Interactive Journal of Medical Research, vol. 13, no. 1, 2024.
[2] J. Yim et al., “Predicting conversion to wet age-related macular degeneration
using deep learning,” Nat. Med., vol. 26, pp. 1250–1255, May 2020.
[3] K. Dvijotham et al., “Enhancing the reliability and accuracy of AI-enabled
diagnosis via complementarity-driven deferral to clinicians,” Nat. Med., vol. 29, no.
7, pp. 1814–1820, July 2023.
[4] O. O’Connor, “Increasing use of artificial intelligence in genomic medicine for
cancer care—The promise and potential pitfalls,” Nature Synthesis, vol. 3, 2025, Art.
no. 20.