Chapter 14 of 17
Generative AI in Healthcare and Agriculture
Examines AI in diagnosis, imaging and clinical documentation, and in crop monitoring, yield prediction and farmer advisories. Numericals cover screening statistics, NDVI and regression, with cases from WHO, ICRISAT and Kisan e-Mitra.
In this chapter
- Part A: Healthcare
- 14.1 Diagnosis assistance
- 14.2 Patient data summarisation and clinical documentation
- 14.3 Limitations and ethics in healthcare AI
- Part B: Agriculture
- 14.4 Crop monitoring
- 14.5 Yield prediction
- 14.6 Farmer advisory chatbots
- 14.7 Worked numericals
- 14.8 Industry and government use cases
- 14.9 Review questions
- Knowledge sources
Part A: Healthcare
14.1 Diagnosis assistance
AI in diagnosis supports, but does not replace, the doctor. It is most mature in reading images.
| Application | What AI does | Example |
|---|---|---|
| Chest X-ray screening | Flags signs of tuberculosis, lung nodules and other findings | Computer-aided detection (CAD) recommended by WHO for TB screening |
| Diabetic retinopathy | Detects eye damage from retinal photographs | Screening in diabetes clinics where eye specialists are scarce |
| Skin conditions | Classifies photos of skin lesions | Triage support in dermatology |
| Pathology | Finds cancer cells in digitised slides | Prioritises slides for a pathologist |
| Clinical question answering | Large language models answer medical questions | Research systems such as Google's Med-PaLM 2 |
Large language models in medicine. Google reported that Med-PaLM scored 67.6% on MedQA, a dataset of US medical licensing exam-style questions, and that Med-PaLM 2 reached 86.5% in 2023. Google itself stresses that passing exam questions is not the same as safe clinical use, and that such models need careful evaluation before use in care (Google Research: Med-PaLM).
The doctor-in-the-loop principle. A clinician remains responsible for every diagnosis and treatment decision. AI output is treated as a second opinion or a prioritisation aid.
14.2 Patient data summarisation and clinical documentation
Doctors spend a large share of their time on records. Generative AI helps by:
- Summarising a patient's history from scattered notes, lab reports and discharge summaries into a one-page view before a consultation.
- Ambient clinical documentation ("AI scribes"): with the patient's consent, the system listens to the consultation and drafts the clinical note for the doctor to review and sign.
- Discharge summaries and referral letters drafted from the record.
- Patient-friendly explanations: turning a technical report into plain language in the patient's own language.
- Coding and billing support: suggesting diagnosis codes for review.
Example prompt (for an approved, privacy-compliant hospital system only)
Summarise the following de-identified patient record for the treating physician.
Sections: Active problems, Current medications with doses, Allergies, Key lab trends (last 3 months), Pending investigations.
Use only the record. Flag any drug interaction or inconsistency for review. Do not suggest a diagnosis.
Record: """..."""
14.3 Limitations and ethics in healthcare AI
| Issue | Why it matters | Safeguard |
|---|---|---|
| Accuracy and hallucination | A wrong dose or invented finding can harm a patient | Clinician review; grounding in the record; no autonomous decisions |
| Bias | Models trained on one population may perform worse on another | Local validation on Indian patient data |
| Privacy | Health data is highly sensitive | De-identification; consent; compliance with the DPDP Act 2023 and hospital policy |
| Accountability | Who is liable for an AI error? | Clear protocols; the clinician signs off |
| Regulation | Software used for diagnosis can be a medical device | In India, CDSCO regulates software as a medical device under the Medical Devices Rules, 2017 |
| Over-reliance | Staff may stop questioning AI output | Training; audit of AI-assisted decisions |
The Indian Council of Medical Research (ICMR) issued Ethical Guidelines for Application of Artificial Intelligence in Biomedical Research and Healthcare in 2023, covering autonomy, safety, data privacy, accountability and fairness.
Part B: Agriculture
14.4 Crop monitoring
Crop monitoring uses images from satellites, drones and phones to track crop health across fields.
- Satellite imagery (for example Sentinel-2 or India's Resourcesat) covers large areas every few days.
- Drones give high-resolution images of a single farm.
- Vegetation indices such as NDVI (Numerical 14.2) measure how green and healthy plants are.
- AI models detect stress, weeds, water shortage and pest attack from these images.
Disease and pest detection. A farmer photographs a leaf with a phone; an image model identifies the likely disease or pest and suggests action. Generative AI adds explanation in the farmer's language and answers follow-up questions.
14.5 Yield prediction
Yield prediction estimates how much a field or region will produce, using:
- weather (rainfall, temperature, humidity);
- soil data (type, nutrients, moisture);
- satellite indices over the season;
- crop variety, sowing date and inputs used;
- past yields.
Models range from simple regression (Numerical 14.3) to machine learning methods such as random forests and deep learning. Governments use regional forecasts for food-security planning; insurers use them for crop insurance; farmers use them to plan sales.
14.6 Farmer advisory chatbots
Generative AI chatbots answer farmers' questions by voice or text in local languages: which seed to use, when to sow, how to treat a pest, how to apply for a scheme. They work best when grounded in verified agricultural extension material, so that they do not invent advice.
14.7 Worked numericals
Numerical 14.1: Why a positive AI screen still needs confirmation (Bayes in practice).
An AI chest X-ray tool has sensitivity 90% and specificity 95%. It screens 10,000 people in a population where 2% have the disease.
| Has disease (200) | No disease (9,800) | |
|---|---|---|
| AI positive | TP = 0.90 × 200 = 180 | FP = 0.05 × 9,800 = 490 |
| AI negative | FN = 20 | TN = 9,310 |
Only about 27% of people flagged positive actually have the disease, though the tool is "90% sensitive". A negative result is very reliable. This is why WHO treats CAD as a screening step followed by a confirmatory diagnostic test.
Numerical 14.2: NDVI for crop health.
Healthy plants reflect much near-infrared (NIR) and absorb red light.
NDVI ranges from −1 to +1; dense healthy vegetation is typically above 0.6. Field A is healthy; Field B shows stress and should be inspected.
Numerical 14.3: Predicting yield with linear regression.
Seasonal rainfall (x, mm) and yield (y, quintals per hectare) for five past seasons:
| x (mm) | 400 | 500 | 600 | 700 | 800 |
|---|---|---|---|---|---|
| y (q/ha) | 12 | 15 | 17 | 20 | 21 |
Means: x̄ = 600, ȳ = 17. The deviations give Sxy = Σ(x − x̄)(y − ȳ) = 2,300 and Sxx = Σ(x − x̄)² = 100,000.
The correlation is r ≈ 0.99, so R² ≈ 0.98: rainfall explains about 98% of the variation in this small, made-up dataset. Real yields depend on many more factors, which is why practical models use many inputs.
Numerical 14.4: Value of a 30% yield gain (illustrative prices).
A farmer grows groundnut on 2 hectares at a usual 1.2 tonnes per hectare. Sowing advice raises yield by 30%, as reported in the ICRISAT pilot. Price is ₹60,000 per tonne.
Numerical 14.5: Doctor time saved by AI summaries (illustrative).
A doctor sees 40 outpatients a day. Reviewing records takes 6 minutes per patient; an AI summary reduces it to 2 minutes.
14.8 Industry and government use cases
- WHO recommendation on AI for TB screening (2021). WHO's updated TB screening guidance, presented in March 2021, recommended computer-aided detection software to read chest X-rays for TB screening in selected populations, followed by confirmatory testing (AuntMinnie report). Indian firms such as Qure.ai make such software.
- Microsoft–ICRISAT sowing app (Andhra Pradesh). In a 2016 pilot in Devanakonda Mandal, Kurnool district, groundnut farmers received sowing-date advice by SMS in Telugu, generated from weather models and historical data. Farmers who followed it reported about 30% higher average yield per hectare (Microsoft India, January 2017).
- Kisan e-Mitra (Government of India). An AI chatbot for the PM-KISAN scheme lets farmers ask questions in their own language. It launched with English, Hindi, Odia, Tamil and Bangla (PIB, 5 December 2023).
- Google Med-PaLM 2 (research). Reached 86.5% on US medical licensing–style questions in 2023, while Google cautioned that more work is needed before safe clinical deployment.
14.9 Review questions
- Analyse the role of AI in diagnosis. Why is the doctor-in-the-loop principle essential?
- List five uses of generative AI in clinical documentation and the safeguards for each.
- An AI tool has 95% sensitivity and 90% specificity; prevalence is 1% among 20,000 people. Find TP, FP, PPV. (Answer: TP 190, FP 1,980, PPV ≈ 8.8%.)
- Compute NDVI for NIR = 0.45, Red = 0.10 and interpret. (Answer: ≈ 0.64, healthy.)
- Using the regression in Numerical 14.3, predict yield for 550 mm of rainfall. (Answer: 15.85 q/ha.)
- Evaluate the Kisan e-Mitra and ICRISAT examples: what made them suited to Indian farmers?
Knowledge sources
- Google Research, Med-PaLM; Singhal, K. et al., "Large language models encode clinical knowledge," Nature 620, 2023.
- WHO, Consolidated guidelines on tuberculosis, Module 2: Screening, 2021.
- ICMR, Ethical Guidelines for Application of Artificial Intelligence in Biomedical Research and Healthcare, 2023.
- Microsoft India, Microsoft and ICRISAT's Intelligent Cloud pilot for Agriculture in Andhra Pradesh increase crop yield for farmers, January 2017.
- Press Information Bureau, Ministry of Agriculture & Farmers Welfare release on PM-KISAN and Kisan e-Mitra, 5 December 2023.
- Rouse, J. W. et al., "Monitoring vegetation systems in the Great Plains with ERTS," NASA SP-351, 1974 (origin of NDVI).
Independent study notes for learning purposes. Not affiliated with any university or course provider. Facts and case-study figures are linked to their original sources. ChatGPT, Gemini, Copilot, Jasper, Synthesia, Runway, Lumen5 and other product names are trademarks of their respective owners.