Reasonable vs. Unreasonable Questions
DeepEvidence needs a complete question sentence to work. A standalone medical term (a disease name, drug name, or lab marker) is not a question — the system can't infer what you actually need from it.
| # | Example input | Rating | Why |
|---|---|---|---|
| 01 | Aspirin | ✗ Poor | Only a single drug name — not a question at all. The AI can't tell whether you want the dose, indications, side effects, or something else. |
| 02 | Can aspirin be used together with a PD-1 drug? | ✓ Good | Names two drugs and states the core question (interaction), so the system can retrieve drug-interaction literature. |
| 03 | The patient is taking aspirin and is scheduled for surgery next week — when should it be stopped? | ✓ Good | Provides the clinical context (pre-op) and specific medication use, with a clear goal (timing of discontinuation), helping return a guideline-based answer. |
| 04 | Diabetes | ✗ Poor | A disease name carries no question by itself. Diagnosis? Treatment? Complications? The system can't guess. |
| 05 | In a patient with type 2 diabetes and heart failure, should an SGLT-2 inhibitor be preferred over metformin? | ✓ Good | Specifies the patient characteristics (type 2 diabetes + heart failure), the drugs compared, and the decision question (preference) — a standard clinical-question format. |
| 06 | Troponin | ✗ Poor | A single lab-marker name. Normal range? Significance of an elevation? How to trend it? You need to state your actual question. |
| 07 | A patient with acute chest pain has a normal initial troponin, but a repeat at 3 hours is elevated — do they need the cath lab immediately? | ✓ Good | Describes the full clinical scenario and the dynamic lab change, and poses a decision question, so the system can accurately match NSTEMI-related guidelines. |
Core principle: a noun is not a question — a sentence is
Each input should include: who (patient characteristics) · what situation (context / medications / labs) · the specific question (what you actually want to know).
Clear vs. Vague Questions
Even a complete sentence can be too vague if it lacks key information. A clear question should include context such as patient age, comorbidities, and specific values, helping the system locate the most relevant evidence.
✗ Vague questions (answers will be generic)
❓ How do I take aspirin?
The system can only give a generic dose range with little clinical utility.
❓ How do I choose an antibiotic?
❓ How is hypertension treated?
✓ Clear questions (the system can give precise, evidence-based answers)
✅ A 65-year-old on aspirin for secondary cardiovascular prevention — what is the recommended dose?
✅ A 65-year-old on aspirin for secondary cardiovascular prevention with chronic kidney disease (eGFR 45 mL/min/1.73m²) — does the dose need adjusting?
✅ Community-acquired pneumonia, inpatient, PORT class III, no penicillin allergy — what is the first-line antibacterial regimen?
PICO framework — the gold standard for building clear clinical questions
P
Patient
Patient characteristics: age, sex,
comorbidities, lab values
I
Intervention
Intervention: drug, procedure,
dose, duration
C
Comparison
Comparison (optional):
another drug or no treatment
O
Outcome
Outcome: survival,
adverse effects, dose adjustment
DeepEvidence ≠ a general LLM
Please don't use DeepEvidence the way you'd use general AI like ChatGPT or Ernie Bot. Their design goals, mechanics, and best-fit scenarios are fundamentally different.
No role-setting needed
DeepEvidence has the evidence-based knowledge system built in; role prompts neither improve nor affect answer quality — they're just unnecessary clutter.
No internet · no web-wide literature search
The system retrieves only from a curated, high-quality medical-literature knowledge base, deliberately avoiding low-quality sources to keep wrong information out of answers.
Won't fabricate references (hallucinations)
Built on a RAG architecture, answers come from real literature snippets, not from the model's invention.
Focused on medicine · not general chat
Non-medical questions reduce efficiency and won't match a dedicated general-purpose model in quality.
| Dimension | General LLM (e.g. ChatGPT) | DeepEvidence |
|---|---|---|
| Knowledge source | Training data (fixed cutoff) | Real-time retrieval from medical-literature databases |
| Role-setting needed | Sometimes helps | ❌ No effect — not needed |
| Web search | Some versions support it | ❌ No internet (anti-hallucination design) |
| Citations | May be fabricated | ✅ Traceable to real literature |
| Clinical diagnostic reasoning | Mediocre, lacks evidence depth | ✅ Purpose-optimized |
| Medical-knowledge quality control | No filtering | ✅ High-quality sources only |
| General writing/chat | ✅ Strong | ❌ Not its purpose |
What can DeepEvidence do?
You can ask directly in all of the following areas — no need to state your role or background.
Clinical & differential diagnosis
Enter symptoms, signs, and lab results to get a systematic differential-diagnosis list and reasoning.
e.g. A 38-year-old woman with joint pain, a malar rash, and positive ANA — what should be in the differential?
Drug use & dosing
Indications, contraindications, dose adjustment in hepatic/renal impairment, and the optimal route and timing.
e.g. For a kidney-transplant patient on tacrolimus, how do I adjust the dose based on blood levels?
Drug interactions
Interaction mechanisms, clinical significance, and management advice when combining drugs.
e.g. When warfarin is combined with fluconazole, how does PT/INR change?
Treatment plans & guidelines
First/second-line treatment per the latest guidelines, and stepwise strategies.
e.g. What is the standard pharmacotherapy for a patient with HFrEF (EF 35%)?
Medical scales & scoring
Interpretation of scores such as CHA₂DS₂-VASc, MELD, APACHE, Wells, and CURB-65.
e.g. Does an atrial-fibrillation patient with a CHA₂DS₂-VASc score of 3 need anticoagulation?
Rare & genetic diseases
Diagnostic criteria, inheritance patterns, genetic-testing advice, and current treatments for rare diseases.
e.g. A child presents with episodic ataxia — which rare diseases should be considered?
Drug side effects & toxicity
Common/rare adverse reactions, toxicity monitoring, and management workflows.
e.g. How is immune-checkpoint-inhibitor-induced pneumonitis graded and managed?
Interpreting tests & investigations
Clinical significance, reference ranges, influencing factors, and next steps for lab markers.
e.g. BNP is markedly elevated but the echocardiogram is normal — what should be considered?
Medical history & basic mechanisms
Disease mechanisms, pathophysiology, and medical-history background.
e.g. What were the main contributions of the Framingham study to cardiovascular disease prevention?
Advanced tips: get more precise answers
Give specific values, not descriptive language
✗ Poor kidney function
✓ eGFR 28 mL/min/1.73m² (CKD stage 4)
List all current medications
✗ The patient is on a few drugs
✓ Currently taking: warfarin 3 mg/d, digoxin 0.125 mg/d, furosemide 20 mg/d
State your decision goal clearly
✗ About this patient…
✓ I need to decide: start anticoagulation now, or wait for repeat testing?
Break complex questions into follow-ups
When a question has multiple decision points, ask the main one first, then follow up after you get the answer.
✓ First: what is the first-line drug? → Then: how is its dose adjusted in renal impairment?
Note special-population status
Pregnancy, lactation, elderly (>80), children, immunosuppression, etc. all affect the recommendation.
✓ The patient is 28 weeks pregnant with hyperthyroidism — which antithyroid drugs can be used?
Note region / resource setting (optional)
If you need Chinese guidelines, or resources are limited at a primary hospital, say so.
✓ At a primary hospital with no CRRT, how should this patient's acute kidney failure be managed?
Common pitfalls when using it
Pitfall 1: treating DeepEvidence like a search engine
A search engine returns a list of links; DeepEvidence returns a synthesized, evidence-based answer. Use complete questions, not keyword searches.
Pitfall 2: expecting it to replace clinical judgment
DeepEvidence provides evidence-based reference; the final clinical decision must be made by a licensed physician considering the specific patient. The system bears no medical liability.
Pitfall 3: thinking longer prompts are better
You don't need a lot of preamble — just include the key patient information plus a clear question. Two or three sentences are enough for a high-quality answer.
Pitfall 4: using it to generate patient-facing copy
Answers are in professional medical language for physicians, not for copying directly to lay patients. For patient-education materials, use general AI to adapt the language.