"""
app.py
======
Mental Health AI — Full Pipeline
Files needed:
mental_xlmr_final/ ← XLM-R model folder
mental_model.h5 ← Survey Keras model
scaler.pkl ← Survey scaler
recommendations.py ← same directory
Install:
pip install streamlit transformers torch tensorflow scikit-learn deep-translator
"""
import sys, os
sys.path.append(os.path.dirname(__file__))
import re, pickle, warnings
import numpy as np
import streamlit as st
import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from deep_translator import GoogleTranslator
sys.path.insert(0, os.path.dirname(__file__))
from recommendations import get_recommendations
warnings.filterwarnings("ignore")
# ── PAGE CONFIG ───────────────────────────────────────────────────────────────
st.set_page_config(page_title="Mental Health AI", page_icon="🧠", layout="wide")
st.markdown("""
""", unsafe_allow_html=True)
# ── CONSTANTS ─────────────────────────────────────────────────────────────────
CLASSES = ["anxiety", "depression", "stress"]
ARABIC_LABELS = {"anxiety": "القلق", "depression": "الاكتئاب", "stress": "الضغط النفسي"}
COLORS = {"anxiety": "#ffa657", "depression": "#79c0ff", "stress": "#56d364"}
SEVERITY_AR = {
"normal": "طبيعي", "mild": "خفيف", "moderate": "متوسط",
"severe": "شديد", "extremely_severe": "شديد جداً", "crisis": "أزمة",
}
SEVERITY_COLORS = {
"normal": "#56d364", "mild": "#e3b341", "moderate": "#ffa657",
"severe": "#f85149", "extremely_severe": "#ff0000", "crisis": "#ff0000",
}
CAUSE_AR = {
"work": "ضغط العمل", "relationships": "العلاقات", "financial": "الضغط المالي",
"academic": "الضغط الأكاديمي", "health": "المخاوف الصحية", "social": "القلق الاجتماعي",
"self_worth": "الثقة بالنفس", "trauma": "الصدمة النفسية", "general": "عام",
}
# ── LOAD MODELS ───────────────────────────────────────────────────────────────
@st.cache_resource
def load_xlmr():
token = st.secrets["HF_TOKEN"]
xlmr_tokenizer = AutoTokenizer.from_pretrained(
"tasneem33355/mental-xlmr", token=token
)
model = AutoModelForSequenceClassification.from_pretrained(
"tasneem33355/mental-xlmr", token=token
)
model.eval()
# Hardcoded — LabelEncoder على ['anxiety','depression','stress'] دايماً بترتّب alphabetically
# 0=anxiety, 1=depression, 2=stress
classes = ["anxiety", "depression", "stress"]
return xlmr_tokenizer, model, classes
@st.cache_resource
def load_survey():
scaler = pickle.load(open(os.path.join(os.path.dirname(__file__), "scaler.pkl"), "rb"))
weights = pickle.load(open(os.path.join(os.path.dirname(__file__), "model_weights.pkl"), "rb"))
def predict(x):
for w in weights:
if len(w) == 2:
x = np.dot(x, w[0]) + w[1]
x = np.maximum(0, x) # ReLU
x = np.exp(x) / np.sum(np.exp(x)) # Softmax
return x
return scaler, predict
xlmr_tokenizer, xlmr_model, le = load_xlmr()
scaler, survey_predict = load_survey()
# ── HELPERS ───────────────────────────────────────────────────────────────────
def clean_text(text):
text = re.sub(r'(.)\1{2,}', r'\1\1', text)
text = re.sub(r'[^\w\s\u0600-\u06FF\[\]]', ' ', text)
return re.sub(r'\s+', ' ', text).strip()
def translate_to_en(text):
try:
return GoogleTranslator(source="auto", target="en").translate(text)
except Exception:
return ""
# ── KEYWORD OVERRIDE ─────────────────────────────────────────────────────────
DEPRESSION_KEYWORDS = [
# عربي فصيح
"اكتئاب", "مكتئب", "مكتئبة", "حزن", "حزين", "حزينة", "يأس", "يائس", "يائسة",
"فراغ", "إحساس بالفراغ", "بلا معنى", "لا معنى", "مالهاش معنى", "بلا هدف",
"لا أمل", "مفيش أمل", "تعبت من الحياة", "زهقت من الحياة",
"مش لاقي معنى", "مش لاقية معنى", "حاسس بالفراغ", "حاسة بالفراغ",
"مفيش طاقة", "مفيش رغبة", "بكاء", "عايز أبكي", "عايزة أبكي",
"وحيد", "وحيدة", "عزلة", "منعزل", "منعزلة",
"إرهاق نفسي", "إرهاق عاطفي", "مش حاسس بحاجة", "مش حاسة بحاجة",
# عامية مصرية وشامية
"زهقت", "تعبت", "مش طايق", "مش طايقة", "نفسيتي وحشة", "نفسيتي في الأرض",
"مش قادر أكمل", "مش قادرة أكمل", "مش عايش", "مش قادر أعيش",
"مش عايز أصحى", "مش عايزة أصحى", "دموع", "بدمع", "قلبي تقيل",
"مش حاسس بنفسي", "مش حاسة بنفسي", "ما بحس بشي", "ما في فايدة",
"مافي امل", "ما في امل", "حياتي خربت", "خسرت كل حاجة",
# إنجليزي
"depressed", "depression", "hopeless", "hopelessness", "empty", "emptiness",
"worthless", "meaningless", "no meaning", "no purpose", "cannot go on",
"cant go on", "no energy", "no motivation", "crying", "feel nothing",
"numb", "isolated", "lonely", "loneliness", "sad", "sadness",
"despair", "grief", "miserable", "broken", "lost all hope",
]
ANXIETY_KEYWORDS = [
# عربي
"قلق", "قلقان", "قلقانة", "خوف", "خايف", "خايفة", "توتر", "متوتر", "متوترة",
"هلع", "مش مرتاح", "مش مرتاحة", "ذعر", "رهاب", "وسواس",
# إنجليزي
"panic", "anxious", "anxiety", "worried", "worry", "fear",
"scared", "nervous", "restless", "tense", "phobia", "ocd",
]
STRESS_KEYWORDS = [
# عربي
"ضغط", "ضغوط", "مضغوط", "مضغوطة", "إجهاد", "مجهد", "مجهدة",
# إنجليزي
"overwhelmed", "stressed", "stress", "burnout", "exhausted", "overloaded",
]
def keyword_boost(text: str, scores: dict) -> dict:
"""
يعوّض الـ stress bias في الموديل عن طريق override قوي
لما تكون كلمات depression أو anxiety أو stress واضحة في النص.
"""
text_lower = text.lower()
dep_hits = sum(1 for kw in DEPRESSION_KEYWORDS if kw.lower() in text_lower)
anx_hits = sum(1 for kw in ANXIETY_KEYWORDS if kw.lower() in text_lower)
str_hits = sum(1 for kw in STRESS_KEYWORDS if kw.lower() in text_lower)
if dep_hits == 0 and anx_hits == 0 and str_hits == 0:
return scores
s = dict(scores)
if dep_hits > 0 and dep_hits >= anx_hits and dep_hits >= str_hits:
# depression كلمات واضحة — override قوي
boost = min(0.55 + dep_hits * 0.10, 0.85)
s["depression"] = boost
remaining = 1.0 - boost
total_rest = s["anxiety"] + s["stress"]
if total_rest > 0:
s["anxiety"] = round(remaining * s["anxiety"] / total_rest, 4)
s["stress"] = round(remaining * s["stress"] / total_rest, 4)
s["depression"] = round(boost, 4)
elif anx_hits > 0 and anx_hits >= dep_hits and anx_hits >= str_hits:
# anxiety كلمات واضحة — override قوي
boost = min(0.55 + anx_hits * 0.10, 0.85)
s["anxiety"] = boost
remaining = 1.0 - boost
total_rest = s["depression"] + s["stress"]
if total_rest > 0:
s["depression"] = round(remaining * s["depression"] / total_rest, 4)
s["stress"] = round(remaining * s["stress"] / total_rest, 4)
s["anxiety"] = round(boost, 4)
elif str_hits > 0 and str_hits >= dep_hits and str_hits >= anx_hits:
# stress كلمات واضحة — override قوي
boost = min(0.55 + str_hits * 0.10, 0.85)
s["stress"] = boost
remaining = 1.0 - boost
total_rest = s["depression"] + s["anxiety"]
if total_rest > 0:
s["depression"] = round(remaining * s["depression"] / total_rest, 4)
s["anxiety"] = round(remaining * s["anxiety"] / total_rest, 4)
s["stress"] = round(boost, 4)
# normalize
total = sum(s.values())
if total > 0:
s = {k: round(v / total, 4) for k, v in s.items()}
return s
def predict_text(text: str) -> dict:
cleaned = clean_text(text)
text_en = translate_to_en(cleaned)
combined = (text_en + " [SEP] " + cleaned) if text_en else cleaned
inputs = xlmr_tokenizer(combined, return_tensors="pt",
truncation=True, max_length=192, padding=True)
with torch.no_grad():
probs = torch.softmax(xlmr_model(**inputs).logits, dim=-1).squeeze().numpy()
raw_scores = {c: round(float(p), 4) for c, p in zip(le, probs)}
# طبّق الـ keyword boost على النص الأصلي + الترجمة
boosted = keyword_boost(text + " " + text_en, raw_scores)
return boosted
def predict_survey(answers: list) -> dict:
data = scaler.transform(np.array(answers).reshape(1, -1))
pred = survey_predict(data)[0]
return {
"depression": round(float(pred[0]), 4),
"anxiety": round(float(pred[1]), 4),
"stress": round(float(pred[2]), 4),
}
def fuse_scores(text_s, survey_s, w_text=0.4, w_survey=0.6):
return {c: round(w_text * text_s[c] + w_survey * survey_s[c], 4) for c in CLASSES}
# ── SURVEY QUESTIONS ─────────────────────────────────────────────────────────
SURVEY_Q = [
("I found it hard to wind down", "وجدت صعوبة في الاسترخاء"),
("I was aware of dryness of my mouth", "لاحظت جفافاً في فمي"),
("I couldn't seem to experience any positive feeling at all", "لم أستطع الشعور بأي مشاعر إيجابية"),
("I experienced breathing difficulty", "أحسست بصعوبة في التنفس"),
("I found it difficult to work up the initiative to do things", "وجدت صعوبة في اتخاذ المبادرة للقيام بالأشياء"),
("I tended to over-react to situations", "كنت أبالغ في ردود أفعالي تجاه المواقف"),
("I experienced trembling", "شعرت بالرعشة"),
("I felt that I was using a lot of nervous energy", "شعرت أنني أستهلك الكثير من الطاقة العصبية"),
("I was worried about situations in which I might panic", "كنت قلقاً من مواقف قد أصاب فيها بالذعر"),
("I felt that I had nothing to look forward to", "شعرت أنه لا يوجد شيء أتطلع إليه"),
("I found myself getting agitated", "وجدت نفسي أشعر بالانفعال"),
("I found it difficult to relax", "وجدت صعوبة في الاسترخاء"),
("I felt down-hearted and blue", "شعرت بالإحباط والكآبة"),
("I was intolerant of anything that kept me from getting on", "كنت غير متسامح مع أي شيء يعيقني"),
("I felt I was close to panic", "شعرت أنني على وشك الذعر"),
("I was unable to become enthusiastic", "لم أستطع أن أتحمس لأي شيء"),
("I felt I wasn't worth much as a person", "شعرت أنني لست شخصاً ذا قيمة"),
("I felt that I was rather touchy", "شعرت أنني متقلب المزاج"),
("I was aware of the action of my heart", "كنت واعياً لنبضات قلبي"),
("I felt scared without any good reason", "شعرت بالخوف دون سبب واضح"),
("I felt that life was meaningless", "شعرت أن الحياة بلا معنى"),
("I found it hard to calm down", "وجدت صعوبة في التهدئة"),
("I felt nervous", "شعرت بالتوتر"),
("I felt sad and depressed", "شعرت بالحزن والاكتئاب"),
("I found myself getting impatient", "وجدت نفسي أشعر بنفاد الصبر"),
("I felt that I was rather emotional", "شعرت أنني عاطفي بشكل مفرط"),
("I felt restless", "شعرت بعدم الهدوء"),
("I had difficulty concentrating", "وجدت صعوبة في التركيز"),
("I felt lonely", "شعرت بالوحدة"),
("I found it difficult to relax", "وجدت صعوبة في الاسترخاء"),
("I felt hopeless", "شعرت باليأس"),
("I felt worried about many things", "كنت قلقاً بشأن أشياء كثيرة"),
("I felt that I had no energy", "شعرت بعدم وجود طاقة"),
("I felt tense", "شعرت بالتوتر والضيق"),
("I felt tired for no reason", "شعرت بالتعب دون سبب"),
("I felt uneasy", "شعرت بعدم الارتياح"),
("I felt worthless", "شعرت بأنني لا قيمة لي"),
("I felt anxious", "شعرت بالقلق"),
("I felt discouraged", "شعرت بالإحباط"),
("I felt stressed", "شعرت بالضغط"),
("I felt overwhelmed", "شعرت بالإرهاق"),
("I felt emotionally exhausted", "شعرت بالإنهاك العاطفي"),
]
# ── UI ────────────────────────────────────────────────────────────────────────
st.title("🧠 Mental Health AI")
st.markdown(
""
"Write how you feel and answer the survey for a complete assessment"
"
اكتب ما تشعر به وأجب على الأسئلة للحصول على تقييم شامل
",
unsafe_allow_html=True,
)
st.markdown("---")
# ── PART 1: TEXT ─────────────────────────────────────────────────────────────
st.markdown("", unsafe_allow_html=True)
st.markdown("### 💬 How are you feeling? / كيف تشعر؟")
st.markdown(
"
"
"Write in any language — Arabic (any dialect), English, or both
"
"اكتب بأي لغة — عربي (أي لهجة)، إنجليزي، أو الاتنين
",
unsafe_allow_html=True,
)
user_text = st.text_area(
label="",
placeholder="e.g. I've been feeling very overwhelmed at work and can't sleep...\nمثال: أنا تعبان جداً من الشغل ومش قادر أنام...",
height=120,
label_visibility="collapsed",
)
st.markdown("
", unsafe_allow_html=True)
st.markdown("### 📋 DASS-42 Survey / استبيان DASS-42")
st.markdown(
"
"
"0 = Never | 1 = Sometimes | 2 = Often | "
"3 = Most of the time | 4 = Always
"
"0 = لم يحدث أبداً | 1 = أحياناً | 2 = كثيراً | 3 = معظم الوقت | 4 = دائماً
",
unsafe_allow_html=True,
)
survey_answers = []
for i in range(0, len(SURVEY_Q), 2):
cols = st.columns(2)
for j, (en, ar) in enumerate(SURVEY_Q[i:i+2]):
with cols[j]:
val = st.slider(
f"{i+j+1}. {en}\n{ar}",
min_value=0, max_value=3, value=0,
key=f"q_{i+j}",
)
survey_answers.append(val)
st.markdown("
"
f"{sev.replace('_',' ').title()} / {SEVERITY_AR.get(sev,'')}
")
col.markdown(f"""
{cls.title()} / {ARABIC_LABELS[cls]}
{pct}%
{badge}
""", unsafe_allow_html=True)
# ── PRIMARY LABEL ─────────────────────────────────────────────────────────
if not rec["suicidal_flag"]:
cause_label = CAUSE_AR.get(rec["cause"], rec["cause"])
st.markdown(
f""
f"Primary: {primary.title()} / {ARABIC_LABELS[primary]}"
f" | Cause detected / السبب المكتشف: "
f"{rec['cause'].replace('_',' ').title()} / {cause_label}
",
unsafe_allow_html=True,
)
# ── CRISIS BOX ────────────────────────────────────────────────────────────
if rec["suicidal_flag"]:
st.markdown("""
🚨 Crisis Support Needed / مطلوب دعم أزمة
""", unsafe_allow_html=True)
# ── SCORE DETAILS ─────────────────────────────────────────────────────────
with st.expander("Show score breakdown / عرض تفاصيل النتائج"):
c1, c2 = st.columns(2)
c1.markdown("**Text model / موديل النص:**")
for cls in CLASSES:
c1.markdown(f"- {cls} / {ARABIC_LABELS[cls]}: **{int(text_scores[cls]*100)}%**")
c2.markdown("**Survey model / موديل السيرفاي:**")
for cls in CLASSES:
c2.markdown(f"- {cls} / {ARABIC_LABELS[cls]}: **{int(survey_scores[cls]*100)}%**")
# ── RECOMMENDATIONS ───────────────────────────────────────────────────────
st.markdown("---")
st.markdown("## 💡 Recommendations / التوصيات")
col_tips, col_res = st.columns(2)
with col_tips:
st.markdown("", unsafe_allow_html=True)
st.markdown("
✅ Practical Tips / نصائح عملية
",
unsafe_allow_html=True)
tips_en = rec.get("tips_en", [])
tips_ar = rec.get("tips_ar", [])
for en, ar in zip(tips_en, tips_ar):
st.markdown(
f"
",
unsafe_allow_html=True,
)
st.markdown("
", unsafe_allow_html=True)
st.markdown("
📚 Resources / موارد مفيدة
",
unsafe_allow_html=True)
res_en = rec.get("resources_en", [])
res_ar = rec.get("resources_ar", [])
for en, ar in zip(res_en, res_ar):
st.markdown(
f"
",
unsafe_allow_html=True,
)
st.markdown("
"
f"🏥 When to seek help / متى تطلب المساعدة:
"
f"{ref_en}
"
f"{ref_ar}"
f"
",
unsafe_allow_html=True,
)
st.markdown(
""
"⚠️ This system is for awareness only and is not a substitute for professional medical diagnosis.
"
"هذا النظام للتوعية فقط وليس بديلاً عن التشخيص الطبي المتخصص.
",
unsafe_allow_html=True,
)