Spaces:
Sleeping
Sleeping
| """ | |
| GLP-1 Response & Safety Stress-Test | |
| A Clinical-AI Validation Demonstration by Dr Adnan Agha | |
| (Consultant Endocrinologist, FRCP London & Glasgow | Clinical-AI Validation Lead) | |
| PURPOSE | |
| ------- | |
| This Space is a DEMONSTRATION of clinical-AI validation methodology, not a clinical tool. | |
| It shows how a tabular ML model can estimate GLP-1 efficacy AND how a *validation lens* | |
| exposes where an efficacy-only model overstates real-world benefit by ignoring adverse | |
| events and the trial-to-real-world gap. | |
| DATA | |
| ---- | |
| The model is trained on SYNTHETIC data generated from published clinical-trial averages | |
| (SURMOUNT-5, SURPASS-2 and related). No real patient data is used. Outputs are illustrative. | |
| NOT A MEDICAL DEVICE. NOT FOR CLINICAL USE. | |
| """ | |
| import numpy as np | |
| import pandas as pd | |
| import matplotlib | |
| matplotlib.use("Agg") | |
| import matplotlib.pyplot as plt | |
| import gradio as gr | |
| from sklearn.ensemble import GradientBoostingRegressor, GradientBoostingClassifier | |
| RNG = np.random.default_rng(42) | |
| # ---------------------------------------------------------------------------- | |
| # Agent parameters grounded in published averages (illustrative magnitudes) | |
| # weight-loss % (magnitude), HbA1c reduction (% points), baseline GI AE rate | |
| # Sources: SURMOUNT-5 (sema -13.7% / tirz -20.2% weight); SURPASS-2 HbA1c | |
| # reductions (-1.86% sema 1mg; -2.01/-2.24/-2.30% tirz 5/10/15mg). | |
| # ---------------------------------------------------------------------------- | |
| AGENTS = { | |
| "Semaglutide 1 mg injectable (Ozempic)": dict(wl=6.0, h1=1.86, gi=0.40), | |
| "Semaglutide 2.4 mg injectable (Wegovy)": dict(wl=13.7, h1=1.86, gi=0.45), | |
| "Oral semaglutide 14 mg (Rybelsus)": dict(wl=3.0, h1=1.4, gi=0.40), | |
| "Tirzepatide 10 mg": dict(wl=19.0, h1=2.24, gi=0.38), | |
| "Tirzepatide 15 mg": dict(wl=20.2, h1=2.30, gi=0.42), | |
| } | |
| AGENT_KEYS = list(AGENTS.keys()) | |
| FEATURES = ["age", "sex", "bmi", "weight", "hba1c", "duration", | |
| "insulin", "su", "metformin", "adherence", | |
| "ag_sema1", "ag_sema", "ag_oral", "ag_t10", "ag_t15"] | |
| # Real-world shrinkage vs trial (real-world HbA1c reductions ~ half of trial; | |
| # weight loss somewhat attenuated). Illustrative, for the validation lens. | |
| RW_H1 = 0.55 | |
| RW_WL = 0.72 | |
| def _agent_onehot(agent): | |
| return [1.0 if agent == k else 0.0 for k in AGENT_KEYS] | |
| def featurize(r): | |
| return [r["age"], r["sex"], r["bmi"], r["weight"], r["hba1c"], r["duration"], | |
| r["insulin"], r["su"], r["metformin"], r["adherence"], | |
| *_agent_onehot(r["agent"])] | |
| def generate_synthetic_data(n=2500): | |
| rows, wl, h1, gi, disc = [], [], [], [], [] | |
| for _ in range(n): | |
| agent = RNG.choice(AGENT_KEYS) | |
| p = AGENTS[agent] | |
| age = float(RNG.normal(54, 11)); age = float(np.clip(age, 25, 80)) | |
| sex = float(RNG.integers(0, 2)) | |
| bmi = float(np.clip(RNG.normal(34, 5), 25, 55)) | |
| weight = float(np.clip(bmi * RNG.normal(2.7, 0.2), 60, 200)) | |
| hba1c = float(np.clip(RNG.normal(8.0, 1.2), 5.8, 12.0)) | |
| duration = float(np.clip(RNG.exponential(7), 0, 30)) | |
| insulin = float(RNG.random() < 0.25) | |
| su = float(RNG.random() < 0.30) | |
| metformin = float(RNG.random() < 0.75) | |
| adherence = float(np.clip(RNG.normal(0.85, 0.12), 0.4, 1.0)) | |
| bmi_factor = float(np.clip(1 + 0.012 * (bmi - 32), 0.80, 1.25)) | |
| age_factor = float(np.clip(1 - 0.003 * (age - 50), 0.85, 1.10)) | |
| weight_loss = -(p["wl"] * bmi_factor * age_factor * adherence + RNG.normal(0, 2.5)) | |
| weight_loss = float(np.clip(weight_loss, -35, 0)) | |
| h1_red = (p["h1"] * (0.6 + 0.12 * (hba1c - 7.0)) | |
| * (1 - 0.018 * duration) | |
| * (0.85 if insulin else 1.0) | |
| * (0.95 if su else 1.0) | |
| * adherence + RNG.normal(0, 0.30)) | |
| h1_red = float(np.clip(h1_red, 0, 4.0)) | |
| gi_prob = float(np.clip(p["gi"] + 0.15 * (1 - adherence) | |
| - 0.002 * (age - 50) + RNG.normal(0, 0.05), 0.05, 0.90)) | |
| gi_event = float(RNG.random() < gi_prob) | |
| disc_prob = float(np.clip(0.10 + 0.30 * (gi_prob - 0.40) | |
| + 0.20 * (1 - adherence) + RNG.normal(0, 0.03), 0.02, 0.50)) | |
| disc_event = float(RNG.random() < disc_prob) | |
| r = dict(age=age, sex=sex, bmi=bmi, weight=weight, hba1c=hba1c, | |
| duration=duration, insulin=insulin, su=su, metformin=metformin, | |
| adherence=adherence, agent=agent) | |
| rows.append(featurize(r)); wl.append(weight_loss); h1.append(h1_red) | |
| gi.append(gi_event); disc.append(disc_event) | |
| df = pd.DataFrame(rows, columns=FEATURES) | |
| df["weight_loss"] = wl; df["h1_red"] = h1; df["gi"] = gi; df["disc"] = disc | |
| return df | |
| def train_models(df): | |
| X = df[FEATURES].values | |
| wl_m = GradientBoostingRegressor(n_estimators=180, max_depth=3, random_state=1).fit(X, df["weight_loss"]) | |
| h1_m = GradientBoostingRegressor(n_estimators=180, max_depth=3, random_state=1).fit(X, df["h1_red"]) | |
| gi_m = GradientBoostingClassifier(n_estimators=160, max_depth=3, random_state=1).fit(X, df["gi"]) | |
| dc_m = GradientBoostingClassifier(n_estimators=160, max_depth=3, random_state=1).fit(X, df["disc"]) | |
| return wl_m, h1_m, gi_m, dc_m | |
| # Train once at startup (synthetic data -> fast, reproducible) | |
| _DF = generate_synthetic_data() | |
| WL_M, H1_M, GI_M, DC_M = train_models(_DF) | |
| def _trajectory_fig(wl_trial, wl_rw): | |
| weeks = np.array([0, 8, 16, 24, 36, 52]) | |
| tau = 16.0 | |
| trial = wl_trial * (1 - np.exp(-weeks / tau)) | |
| rw = wl_rw * (1 - np.exp(-weeks / tau)) | |
| fig, ax = plt.subplots(figsize=(5.2, 3.2), dpi=120) | |
| ax.plot(weeks, trial, marker="o", linewidth=2.2, label="Trial-style estimate") | |
| ax.plot(weeks, rw, marker="s", linewidth=2.2, linestyle="--", label="Real-world-adjusted") | |
| ax.fill_between(weeks, trial, rw, alpha=0.12) | |
| ax.set_xlabel("Weeks since initiation"); ax.set_ylabel("Weight change (%)") | |
| ax.set_title("Illustrative weight-loss trajectory (synthetic)", fontsize=10) | |
| ax.axhline(0, color="#999", linewidth=0.8) | |
| ax.legend(fontsize=8, loc="lower left"); ax.grid(alpha=0.25) | |
| fig.tight_layout() | |
| return fig | |
| def run_prediction(age, sex, bmi, weight, hba1c, duration, insulin, su, metformin, adherence, agent): | |
| sex_v = 1.0 if sex == "Male" else 0.0 | |
| r = dict(age=age, sex=sex_v, bmi=bmi, weight=weight, hba1c=hba1c, duration=duration, | |
| insulin=float(bool(insulin)), su=float(bool(su)), metformin=float(bool(metformin)), | |
| adherence=adherence / 100.0, agent=agent) | |
| x = np.array([featurize(r)]) | |
| wl_trial = float(WL_M.predict(x)[0]) | |
| h1_trial = float(H1_M.predict(x)[0]) | |
| gi_prob = float(GI_M.predict_proba(x)[0][1]) | |
| disc_prob = float(DC_M.predict_proba(x)[0][1]) | |
| # Validation lens: real-world adjustment + discontinuation drag | |
| h1_rw = h1_trial * RW_H1 * (1 - 0.5 * disc_prob) | |
| wl_rw = wl_trial * RW_WL * (1 - 0.4 * disc_prob) | |
| h1_gap = h1_trial - h1_rw | |
| flags = [] | |
| if gi_prob > 0.50: | |
| flags.append("**Elevated GI adverse-event risk** — tolerability and titration strategy materially affect real-world results.") | |
| if disc_prob > 0.20: | |
| flags.append("**Elevated discontinuation risk** — an efficacy-only model overstates *population* benefit because a meaningful fraction stop therapy.") | |
| if h1_gap > 0.7: | |
| flags.append(f"**Large trial-to-real-world gap (~{h1_gap:.1f}% HbA1c)** — headline efficacy should be discounted for this profile.") | |
| if not flags: | |
| flags.append("No high-risk flags for this profile, but real-world attenuation still applies.") | |
| md = f""" | |
| ### ⚠️ Demonstration only — not for clinical use | |
| *Synthetic model. Illustrative outputs. See **Methodology** below.* | |
| #### Efficacy — trial-style estimate | |
| - **HbA1c reduction:** −{h1_trial:.1f}% points | |
| - **Weight change (52 wk):** {wl_trial:.1f}% | |
| #### Safety — pharmacovigilance signals | |
| - **GI adverse-event risk:** {gi_prob*100:.0f}% | |
| - **Treatment-discontinuation risk:** {disc_prob*100:.0f}% | |
| #### 🔍 Validation lens — what an efficacy-only model misses | |
| | | Trial-style | Real-world-adjusted | | |
| |---|---|---| | |
| | HbA1c reduction | −{h1_trial:.1f}% | **−{h1_rw:.1f}%** | | |
| | Weight change | {wl_trial:.1f}% | **{wl_rw:.1f}%** | | |
| {chr(10).join('- ' + f for f in flags)} | |
| > The point of this demo: a model can be highly *accurate* on efficacy and still be *unsafe to deploy* if it ignores adverse events, discontinuation, and the trial-to-real-world gap. Validating for that is the work. | |
| """ | |
| return md, _trajectory_fig(wl_trial, wl_rw) | |
| DISCLAIMER = """ | |
| # 🩺 GLP-1 Response & Safety Stress-Test | |
| ### A clinical-AI validation demonstration — Dr Adnan Agha, Consultant Endocrinologist (FRCP London & Glasgow) | |
| > **⚠️ DEMONSTRATION ONLY — NOT FOR CLINICAL USE.** This is an educational demonstration of clinical-AI | |
| > *validation methodology*. It is **not** a medical device, is **not** validated for patient care, and must | |
| > **not** be used to make treatment decisions. The model is trained on **synthetic data generated from | |
| > published trial averages** — not on real patient records. All outputs are illustrative. | |
| """ | |
| METHODOLOGY = """ | |
| ### Methodology & data provenance | |
| - **Synthetic data only.** Training data is generated from published clinical-trial average effect sizes | |
| (e.g., SURMOUNT-5 weight loss: tirzepatide −20.2% vs semaglutide −13.7%; SURPASS-2 HbA1c reductions | |
| −1.86% to −2.30%), with plausible predictor relationships and added noise. **No real patient data is used.** | |
| - **Model.** Gradient-boosted trees estimate efficacy (HbA1c reduction, weight change) and adverse-event / | |
| discontinuation risk — the tabular approach well suited to structured clinical prediction. | |
| - **The validation lens.** A real-world shrinkage factor and a discontinuation drag are applied to the | |
| trial-style estimate to show how efficacy-only models overstate population benefit. This is the core | |
| message: validation must stress-test for safety and real-world behaviour, not just headline accuracy. | |
| - **Predictors used (illustrative):** baseline BMI and weight (strongest for weight loss); baseline HbA1c, | |
| diabetes duration, and concurrent insulin/sulfonylurea (for glycemic response); adherence; agent. | |
| *For a real validated tool, this would require IRB/ethics-approved real-world data, prospective validation, | |
| and regulatory assessment — services my practice provides under the DoH/THREC framework.* | |
| """ | |
| ABOUT = """ | |
| ### About this demonstration | |
| Built by **Dr Adnan Agha** — Consultant Endocrinologist (FRCP London & Glasgow; dual UK CCT) and clinical-AI | |
| developer. I build and independently **validate** clinical-AI systems, with a focus on **GLP-1 / cardiometabolic | |
| real-world evidence** and **medical-AI safety**. My fine-tuned model *Pentabrid* ranked **#3 globally on the | |
| MedXpertQA leaderboard (ICML 2025)**. | |
| **This tool exists to make one point tangible:** most predictive models in this space optimise for efficacy and | |
| quietly ignore adverse events and the trial-to-real-world gap. Closing that gap — validating clinical AI for | |
| *safety*, not just accuracy — is what I do for health-tech teams and medical-affairs groups. | |
| **Independent clinical-AI validation · GLP-1 / cardiometabolic RWE · advisory & evaluation.** | |
| Contact: adnanagha@uaeu.ac.ae · [LinkedIn URL] · Al Ain / Abu Dhabi, UAE | |
| *(External engagements undertaken in line with institutional approval.)* | |
| """ | |
| with gr.Blocks(title="GLP-1 Response & Safety Stress-Test") as demo: | |
| gr.Markdown(DISCLAIMER) | |
| with gr.Row(): | |
| with gr.Column(scale=1): | |
| gr.Markdown("#### Patient parameters *(hypothetical)*") | |
| agent = gr.Dropdown(AGENT_KEYS, value="Tirzepatide 15 mg", label="GLP-1 agent") | |
| age = gr.Slider(25, 80, value=55, step=1, label="Age (years)") | |
| sex = gr.Dropdown(["Female", "Male"], value="Female", label="Sex") | |
| bmi = gr.Slider(25, 55, value=34, step=0.5, label="Baseline BMI (kg/m²)") | |
| weight = gr.Slider(60, 200, value=95, step=1, label="Baseline weight (kg)") | |
| hba1c = gr.Slider(5.8, 12.0, value=8.2, step=0.1, label="Baseline HbA1c (%)") | |
| duration = gr.Slider(0, 30, value=6, step=1, label="Diabetes duration (years)") | |
| adherence = gr.Slider(40, 100, value=85, step=1, label="Expected adherence (%)") | |
| with gr.Row(): | |
| insulin = gr.Checkbox(label="On insulin") | |
| su = gr.Checkbox(label="On sulfonylurea") | |
| metformin = gr.Checkbox(value=True, label="On metformin") | |
| btn = gr.Button("Run stress-test", variant="primary") | |
| with gr.Column(scale=1): | |
| out_md = gr.Markdown() | |
| out_plot = gr.Plot() | |
| with gr.Accordion("Methodology & data provenance", open=False): | |
| gr.Markdown(METHODOLOGY) | |
| with gr.Accordion("About / contact", open=False): | |
| gr.Markdown(ABOUT) | |
| btn.click(run_prediction, | |
| [age, sex, bmi, weight, hba1c, duration, insulin, su, metformin, adherence, agent], | |
| [out_md, out_plot]) | |
| demo.load(run_prediction, | |
| [age, sex, bmi, weight, hba1c, duration, insulin, su, metformin, adherence, agent], | |
| [out_md, out_plot]) | |
| if __name__ == "__main__": | |
| demo.launch() | |